Roof distributed photovoltaic installation potential assessment method based on deep learning

Abstract

The invention belongs to the technical field of electric power, and discloses a roof distributed photovoltaic installation potential evaluation method based on deep learning, which is particularly suitable for being applied to an urban environment. The problem that in the prior art, the prediction precision of the power consumption demand of a user is insufficient is solved, and a composite prediction model combining variational mode decomposition and a Crossform network is disclosed. According to the model, firstly, VMD is used for preprocessing original power utilization data, modal components of different frequencies are effectively separated out, noise interference is reduced, and data quality is improved; and further analyzing and capturing a long-term dependency relationship in the time sequence through a Crossform network, thereby realizing high-precision prediction of the power consumption demand. The method improves the accuracy and practicability of evaluation, can be widely applied to different urban environments, and provides scientific basis and technical support for urban energy planning. The invention further discloses a comprehensive photovoltaic installation potential evaluation framework.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power, and particularly relates to a method for evaluating the installation potential of rooftop distributed photovoltaic based on deep learning. Background Art

[0002] Clean energy, especially solar energy, has become an important direction for energy transformation due to its renewable and environmentally friendly characteristics. As a new type of solar energy utilization method, the Distributed Photovoltaic (DPV) system plays an increasingly important role in urban energy supply due to its high flexibility, small footprint, and easy integration into urban buildings. In order to better promote and optimize the DPV system, accurately evaluating its installation potential has become a key issue.

[0003] When evaluating the installation potential of rooftop distributed photovoltaic systems, accurately predicting the electricity demand of users is crucial. This is because by understanding the specific electricity consumption needs of users, the installed capacity of the photovoltaic system can be determined more scientifically and reasonably, thus ensuring the balance between supply and demand and maximizing the economic and environmental benefits of the system. Electricity demand forecasting is affected by many external factors, such as weather conditions, seasonal changes, and socio-economic activities. These factors make the electricity demand show obvious non-linear and time-varying characteristics, increasing the difficulty of forecasting.

[0004] Currently, the mainstream electricity demand forecasting methods are mainly divided into two types. One is the electricity demand forecasting method mainly based on signal analysis, such as empirical mode decomposition, variational mode decomposition, and singular value decomposition. This method mainly relies on manually extracting the characteristic values of the load signal to forecast the electricity demand. The other is the deep learning method mainly based on data-driven, which does not require manual participation but the results of electricity demand forecasting are not accurate enough. There are technical problems in the prior art with low accuracy and reliability in electricity consumption demand forecasting.

[0005] CN118412865A discloses a method for predicting the power generation of a photovoltaic power station based on complex weather, which relates to the technical field of photovoltaic power station power generation, and includes the following steps: obtaining historical power generation data, where the historical power generation data includes historical power generation time series signals and weather type time series signals; cutting the historical power generation data according to the weather type time series signals to obtain several signal sets belonging to the same weather type; performing variational mode decomposition on the signal sets to obtain several discontinuous mode component signals with different frequencies; performing equidistant cutting and self-comparative complementation on the discontinuous mode component signals to obtain smooth mode component signals; analyzing the power generation power law under the corresponding weather type based on the smooth mode component signals in combination with the corresponding weather type, and predicting the power generation power of the photovoltaic power station according to the power generation power law, which realizes the prediction of the power generation power of the photovoltaic power station under complex weather conditions.

[0006] CN118627929A discloses a power load prediction system based on artificial intelligence, including a data acquisition module, a feature processing module, a prediction driving module, a prediction application module, and a monitoring and feedback module. It uses a dual-path feature extraction idea of overlapping feature extraction and overlapping feature decomposition for overlapping feature extraction and decomposed feature extraction, optimizing the overall performance of the system; uses a time-varying filter to improve the traditional empirical mode decomposition, optimizing the performance of feature extraction; uses a wavelet transform improvement method combining variational mode decomposition, smoothing function, and marginal spectrum analysis to further improve the quality of features; uses a multi-path recurrent bidirectional long short-term memory network combined with convolutional time feature extraction for classified load prediction, optimizing the overall architecture of the prediction model.

[0007] The applicant believes that there is still room for improvement, and it needs to be further improved in terms of the accuracy and reliability of electricity demand prediction. Summary of the Invention

[0008] In view of this, it is necessary to provide a method for evaluating the installation potential of rooftop distributed photovoltaics based on deep learning to solve the technical problems of low accuracy and reliability in evaluating the installation potential of rooftop distributed photovoltaics in the prior art.

[0009] To achieve the above object, the present invention provides a method for evaluating the installation potential of rooftop distributed photovoltaics based on deep learning, including:

[0010] Obtaining the original electricity demand data, decomposing the original electricity demand data based on the VMD algorithm to obtain multiple sub-modal data;

[0011] Based on the original electricity demand data and the multiple sub-modal data, obtaining reconstructed data, and dividing the reconstructed data to obtain a training set, a validation set, and a test set;

[0012] Obtain the Crossformer network model, input the training set and the validation set into the Crossformer network model for iterative training to obtain an electricity demand prediction model;

[0013] Input the test set into the electricity demand prediction model to obtain an electricity demand prediction result.

[0014] In a possible implementation manner, decomposing the original electricity demand data based on the VMD algorithm to obtain multiple sub-modal data, including:

[0015] Obtain the decomposition layer number, and sequentially decompose the original electricity demand data according to the decomposition layer number based on the VMD algorithm to obtain multiple sub-modal data.

[0016] In a possible implementation manner, the obtaining of the decomposition layer number includes:

[0017] Adaptive adjustment of the decomposition layer number based on the correlation coefficient to obtain the optimal decomposition layer number.

[0018] In a possible implementation manner, the electricity demand prediction model includes:

[0019] Dimension segmentation embedding, two-stage attention layer and feature fusion module;

[0020] Input the test set into the electricity demand prediction model, segment the test set through the dimension segmentation embedding to obtain multiple segmented data, obtain feature vectors, and embed the multiple segmented data into the feature vectors to obtain multiple two-dimensional vector arrays. Capture features of the two-dimensional vector arrays through the two-stage attention layer to obtain captured features, and merge the multiple two-dimensional vector arrays based on the captured features through the feature fusion module to obtain multiple merged vectors, and obtain an electricity demand prediction result based on the multiple merged vectors.

[0021] In a possible implementation manner, the feature fusion module includes: a three-layer encoder-decoder unit;

[0022] Merge and predict the multiple two-dimensional vector arrays based on the captured features through the three-layer encoder-decoder unit in the feature fusion module to obtain multiple merged vectors.

[0023] In a possible implementation manner, the three-layer encoder-decoder unit includes: a first-layer encoder-decoder unit;

[0024] The first-layer encoder-decoder unit includes: an encoder and a decoder; the output end of the first-layer encoder is connected to the input end of the first-layer decoder;

[0025] The first - layer encoder in the first - layer encoder - decoder unit of the three - layer encoder - decoder unit merges the multiple two - dimensional vector data based on the captured features to obtain a feature representation of the first - layer encoder, and predicts the feature representation of the first - layer encoder based on the first - layer decoder in the first - layer encoder - decoder unit of the three - layer encoder - decoder unit to obtain a first - layer merged vector.

[0026] In a possible implementation, obtaining the electricity demand prediction result based on multiple merged vectors includes:

[0027] Performing an addition process on the multiple merged vectors to obtain the electricity demand prediction result.

[0028] To achieve the efficient operation of the power system and the effective utilization of renewable energy, it is particularly important to integrate load forecasting and distributed photovoltaic installation potential assessment. By using a deep - learning model for high - precision load forecasting and evaluating the photovoltaic installation potential by combining geographic information and meteorological data, etc., the power demand and supply can be accurately matched, the layout and scale of the photovoltaic system can be optimized, thereby improving energy utilization efficiency, reducing environmental pollution, and promoting sustainable development. Specifically, high - precision load forecasting can help power companies make power generation plans in advance and avoid power shortages or surpluses caused by the imbalance between supply and demand; while the assessment of photovoltaic installation potential can determine the potential capacity suitable for installing photovoltaic systems in a certain area, providing a scientific basis for the government to formulate relevant policies and support measures. By integrating load forecasting and photovoltaic installation potential assessment, not only can the flexibility and reliability of the power system be improved, but also the application and development of clean energy can be effectively promoted, providing strong support for achieving the goal of green and low - carbon. In addition, the integrated model can also be used to evaluate the economic feasibility of distributed photovoltaic projects. By predicting the future load growth trend and combining factors such as the investment cost, operation and maintenance costs of the photovoltaic system, and possible subsidy policies, etc., the return on investment of the project can be comprehensively analyzed, providing decision - making support for investors. At the same time, this integration also helps to optimize the operation strategy of the power system, improve energy utilization efficiency, reduce carbon emissions, promote the development of green energy, and ultimately achieve a win - win situation in economic and environmental benefits. Brief Description of the Drawings

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those skilled in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0030] Figure 1 It is the model prediction process provided by the present invention.

[0031] Figure 2 It is the waveform diagram of the original power demand data provided by the present invention.

[0032] Figure 3 It is the waveform diagram of the power demand decomposition result provided by the present invention.

[0033] Figure 4 It is the schematic diagram of the architecture of the hierarchical encoder-decoder provided by the present invention.

[0034] Figure 5 It is the flowchart of signal processing based on the improved VMD algorithm provided by the present invention.

[0035] Figure 6 It is the schematic diagram of model comparison provided by the present invention.

[0036] Figure 7 It is the framework structure of the evaluation model provided by the present invention.

[0037] Figure 8 It is the chart of the measurement standard for the correlation of the present invention.

[0038] Figure 9 It is the correlation coefficient table between each modal component and the original data of the present invention.

[0039] Figure 10 It is the experimental data of model comparison of the present invention. Detailed implementation manners

[0040] [[ID=4o]]Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts fall within the protection scope of the present invention. In the description of the embodiments of the present application, unless otherwise specified, the meaning of "a plurality" is two or more.

[0041] The present invention discloses a method for evaluating the installation potential of rooftop distributed photovoltaic based on deep learning, which will be described separately below.

[0042] Figure 1 It is a schematic flowchart of an embodiment of the deep learning model provided by the present invention, including:

[0043] S101. Obtain the original power demand data, and decompose the original power demand data based on the VMD algorithm to obtain a plurality of sub-modal data;

[0044] S102. Obtain reconstructed data based on the original power demand data and the multiple sub-modal data, and perform data partitioning on the reconstructed data to obtain a training set, a validation set, and a test set;

[0045] S103. Obtain a Crossformer network model, input the training set and the validation set into the Crossformer network model for iterative training to obtain a deep learning model for power demand prediction;

[0046] S104. Input the test set into the power demand prediction model to obtain a power demand prediction result.

[0047] It can be understood that in step S101, the original power demand data includes time and power consumption demand.

[0048] Specifically, in this embodiment, an open power demand dataset is adopted, which is highly reliable and has a low data error and missing rate.

[0049] In step S101, the VDM algorithm is an adaptive signal decomposition method that can automatically adjust the decomposition parameters according to the characteristics of the signal, thereby obtaining more accurate modal components. The original signal to be processed is decomposed into K variational modal components with different center frequency bandwidths, and it is ensured that the sum of the estimated bandwidths of all modal components is minimized. The specific variational constraint expression can be represented as:

[0050] By introducing a quadratic penalty factor and a Lagrange multiplier the constrained problem is transformed into an unconstrained variational problem. Among them, is used to balance the trade-off between the smoothness of the signal and the exact reconstruction, is used to impose constraint conditions.

[0051] ;

[0052] The alternating direction multiplier method is adopted to continuously iteratively update and and and search for the saddle point of the augmented Lagrangian. The iterative process is as follows:

[0053] ;

[0054] ;

[0055] ;

[0056] In the formula, , and are respectively , , the Fourier transform of, is the number of iterations, is the iteration step size. Combining the processes and stopping conditions of the above three equations:

[0057] ;

[0058] In the formula, is the convergence accuracy. If the stopping condition in this formula is not satisfied, is increased to , and it stops until the stopping condition is satisfied or the maximum number of iterations is reached. Finally, modal components with independent center frequencies and finite bandwidths are obtained.

[0059] It has been found through research that the number of modes K needs to be set manually and has a great influence on the signal decomposition effect. To solve this problem, the present invention uses the correlation coefficient between the original signal and each component of VMD to adaptively adjust the VMD parameter K. The Pearson correlation coefficient can measure the linear correlation degree between each modal component and the original signal, and it has a wide range of applications in fields such as data analysis and fault diagnosis. Its calculation formula is:

[0060] ;

[0061] In the formula, the variables and represent the original load data and different modal components respectively, is the average value of the elements in, and N is the number of samples. Figure 8 lists the measurement criteria for correlation.

[0062] After performing VMD decomposition on the original data, K modal components are obtained, and the correlation coefficient between each modal component and the original signal is calculated to measure the corresponding correlation degree. When the absolute value of the correlation coefficient is less than 0.2, it can be considered that there is almost no correlation between the two variables; when the absolute value of the correlation coefficient is higher than 0.2, it indicates that each modal component is an effective component and the signal is still in an under-decomposed state, and the value of K needs to be increased. When the absolute value of the minimum correlation coefficient is first lower than 0.2, it is considered that the important information in the signal has been fully decomposed, and the K at this time is the optimal decomposition layer.

[0063] It can be further understood that in step S102, the data is reconstructed, and the data set is divided into a training set, a validation set, and a test set;

[0064] This step specifically includes:

[0065] The decomposed sub-modal data and the original data are combined to form reconstructed data, and the reconstructed data is divided into a training set, a validation set, and a test set according to a ratio of 6:1:3.

[0066] In step S103, the training set and the validation set are input into the established Crossformer network model for training to obtain a load prediction model;

[0067] The process of this step is as follows:

[0068] The mean squared error is used as the loss function and the Adam optimizer to train the model.

[0069] In step S104, the test set is input into the electricity demand prediction model to obtain an electricity demand prediction result.

[0070] The process of this step is as follows:

[0071] The test set is input into the electricity demand prediction model for electricity demand prediction to obtain a prediction evaluation and a prediction result.

[0072] The present invention also conducts experimental verification and performance analysis on the provided electricity demand prediction model.

[0073] Figure 2 The waveform diagram of the original electricity demand data for an embodiment of the electricity demand prediction method based on deep learning provided by the present invention includes:

[0074] Obtain the decomposition level, and based on the VMD algorithm, decompose the original electricity demand data according to the decomposition level in sequence to obtain multiple sub-modal data.

[0075] It should be noted that in order to verify the effectiveness of the method of the present invention, the dataset used in this experiment is only the publicly available demand dataset. Figure 3 The schematic diagram of the VMD (Variational Mode Decomposition) result for an embodiment of the power load prediction method based on deep learning provided by the present invention.

[0076] Figure 5 The flowchart of signal processing based on the improved VMD algorithm provided by the present invention specifically includes:

[0077] The dataset is decomposed by an improved VMD method for the original data. The correlation coefficients between each component and the original data under different numbers of modes are shown in Table 2. For both datasets, the correlation coefficient first drops below 0.2 when K = 4. Therefore, the optimal decomposition level is 4. Figure 9 The table of the correlation coefficients between each modal component and the original data.

[0078] According to the VMD parameter setting method, set the number of modes K to 4, and the quadratic penalty factor = 2000, and other parameters take the VMD default values. The modal components obtained by VMD decomposition. The modal component data and the original load data in the dataset are respectively composed into reconstructed data and input into the Crossformer prediction network.

[0079] In a possible implementation, the obtaining of the decomposition layer number includes:

[0080] Based on the correlation coefficient, adaptively adjust the decomposition layer number to obtain the optimal decomposition layer number.

[0081] Referring to Figure 4 , it can be understood that, first, initialize the decomposition layer number K, that is, K = 2. Secondly, perform VMD decomposition on the electricity demand data and obtain K modal components. Then, calculate the correlation coefficient p between the original data and each modal decomposition. Finally, judge whether the smallest is lower than 0.2. If not, then , and perform VMD decomposition and calculate the correlation coefficient again; if so, the current decomposition layer number is the optimal decomposition layer number K. After determining the decomposition layer number K, VMD decomposes the original load data into multiple sub-modes.

[0082] It needs to be further understood that the steps for establishing the deep learning model are as follows:

[0083] In the design process of the Crossformer electricity demand prediction model, important hyperparameters include the segment length, the number of attention heads, the number of routers, and the batch size, which directly affect the network prediction effect. The specific values are 8, 6, 8, and 4 respectively. During the experiment, the number of iterations is 10, the learning rate is 1e-4, the number of encoder layers is 3, and the number of decoder layers is 4. When finding the optimal value of one parameter, the remaining parameters remain unchanged. The optimal parameters are determined one by one through experiments below. Use the mean square error as the loss function and use the Adam (adaptive moment estimation) optimizer for training. Using the parameters finally determined by the experiment, the average time per epoch in the dataset is 670.413s.

[0084] Figure 4 It is a schematic diagram of the hierarchical encoder-decoder architecture of an embodiment of the electricity demand prediction method based on deep learning provided by the present invention, including: three-layer encoder-decoder units;

[0085] Through the three-layer encoder-decoder units in the feature fusion module, merge and predict the multiple two-dimensional vector arrays based on the captured features to obtain multiple merged vectors.

[0086] It is understandable that the feature fusion module has an architecture of a hierarchical encoder-decoder with 3 encoder layers. HED is constructed by DSW embedding, TSA layer and segment merging. The encoder uses the TSA layer and segment merging to capture dependencies at different scales: the vectors in the upper layer cover a longer range, resulting in dependencies at a coarser scale. Exploring different scales, the decoder makes a final prediction by making predictions at each scale and adding them up.

[0087] In some embodiments, the three-layer encoder-decoder unit includes: a first-layer encoder-decoder unit;

[0088] The first-layer encoder-decoder unit includes: an encoder and a decoder; the output end of the first-layer encoder is connected to the input end of the first-layer decoder;

[0089] Based on the captured features, the first-layer encoder in the first-layer encoder-decoder unit of the three-layer encoder-decoder unit merges the multiple two-dimensional vector data to obtain a feature representation of the first-layer encoder, and based on the first-layer decoder in the first-layer encoder-decoder unit of the three-layer encoder-decoder unit, makes a prediction on the feature representation of the first-layer encoder to obtain a first-layer merged vector.

[0090] It is understandable that in the N-layer encoder, except for the first layer, each layer merges every two adjacent vectors in the time domain to obtain a representation at a coarser level, and then applies the TSA layer to capture the dependencies at this scale.

[0091] In the decoder, the feature arrays output by the encoder are obtained, and layers are used to make predictions in the decoder. A linear projection is applied to the output of each layer to obtain the prediction of this layer. The predictions of each layer are added up to obtain the final prediction.

[0092] In some embodiments, obtaining the electricity demand prediction result based on the multiple merged vectors includes:

[0093] Performing an addition process on the multiple merged vectors to obtain the electricity demand prediction result.

[0094] It is understandable that the predictions output by the three-layer encoder-decoder unit are added up to obtain the final prediction.

[0095] Figure 6 Schematic diagram of a model comparison experiment for an embodiment of the electricity demand prediction method based on deep learning provided by the present invention, including:

[0096] After training the electricity demand prediction model, the test set is brought into the designed model for load prediction. To verify the feasibility and superiority of the proposed method, the present invention reflects the role of each module through model comparison experiments, such as Figure 10 shown.

[0097] It can be understood that through the model comparison experiment, it can be seen that the addition of each module of the present invention significantly improves the prediction rate, and MAE, MAPE, and RMSE are the lowest on the data set. Compared with the other three prediction methods, the prediction method using VMD-Crossformer has the lowest MAE, MAPE, and RMSE, that is, the VMD-Crossformer model has higher prediction accuracy. Specifically, compared with EMD-PSO-GRU, VMD-SG-LSTM, and VMD-Pyraformer-Adan in the model of the present application, MAE is reduced by 2.201MW to 11.35MW; MAPE is reduced by 0.067% to 0.302%; RMSE is reduced by 3.468MW to 15.639MW.

[0098] Select 120 sample points and visualize the prediction results of the comparison experiment as Figure 6 shown. It can be intuitively observed that the results of the prediction model in this paper are more fitting to the real data curve. For complex power load data, the model in the present application can capture more feature information when the data suddenly increases or decreases compared with other models, thus obtaining higher accuracy. It can be seen from the figure that there is mode mixing in EMD decomposition, which makes GRU unable to effectively learn important features when processing the decomposed signal, resulting in low prediction accuracy of EMD-PSO-GRU. The performance of LSTM has a certain delay, resulting in a decrease in the prediction result, making the prediction accuracy of VMD-SG-LSTM not the highest. The Transformer variant can capture global information well through the self-attention mechanism, making the prediction accuracy of the VMD-Pyraformer-Adan and VMD-Crossformer models better, and the Crossformer network captures cross-dimensional dependencies more directly than the Pyraformer network, making the prediction accuracy better.

[0099] Figure 7 The framework structure of the evaluation model provided by the present invention is as follows: The steps for evaluating the installed capacity potential of rooftop distributed photovoltaic are as follows:

[0100] First, the Pearson correlation coefficient (PCC) is used to determine the optimal decomposition level of VMD. The sub-modal data obtained by decomposing the electricity demand data through VMD is reconstructed with the original data, and the reconstructed data is used for prediction through a deep learning model. This deep learning model adopts DSW embedding and focuses on two stages: cross-time and cross-dimension, making the prediction results more accurate. Based on the high-precision prediction of electricity demand by the deep learning model, we analyze the temporal distribution characteristics of electricity demand. Special attention is paid to the peak demand period and the trough demand period, and the information in these periods helps to determine the optimal time and location for photovoltaic installation to maximize the satisfaction of peak demand.

[0101] Secondly, when considering the relationship between the evaluation of the rooftop distributed photovoltaic installation potential and the supply of electricity demand, the goal is to ensure that the electric energy generated by the photovoltaic system can meet the electricity demand of users without causing excess waste. When calculating the area Si of photovoltaic panels to be installed, first, based on the electricity demand prediction results and combined with the specific conditions of the rooftop (such as physical factors like area, orientation, and shading conditions), the total area Si of photovoltaic panels required to meet the electricity demand of users is calculated. This process needs to comprehensively consider factors such as the spatial limitations of the rooftop, lighting conditions, and building types to ensure that the calculation results are both scientific and reasonable and can be practically implemented.

[0102] Finally, determine the actual feasible installation area of photovoltaic panels Min{Si, Sa}. Here, Sa represents the actual available area of the rooftop. By comparing Si and Sa, the smaller value of the two is selected as the final installation area of photovoltaic panels. This is done to ensure that the installation of the photovoltaic system does not exceed the actual available space of the rooftop and can meet the electricity demand of users as much as possible, thereby achieving the efficient use of resources and the optimal configuration of the system.

[0103] The present invention proposes a composite prediction model of VMD and Crossformer network for electricity demand prediction, and on this basis, constructs an evaluation framework for photovoltaic installation potential. This framework comprehensively considers factors such as the available area of the rooftop, sunlight conditions, building types, and geographical locations, and can accurately evaluate the installation potential of rooftop distributed photovoltaic systems. Experimental results show that the composite prediction model is significantly superior to traditional methods in terms of prediction accuracy, and the evaluation framework shows high practicality and promotion value in practical applications, providing strong support for the green transformation of the urban energy structure.

[0104] Through experimental research, in terms of the superiority of the composite prediction model, through experimental comparison, it is found that the composite prediction model combining Variational Mode Decomposition (VMD) and Crossformer network shows higher accuracy in electricity demand prediction. The VMD preprocessing effectively reduces data noise and improves data quality, while the Crossformer network successfully captures the long-term dependencies in the time series, significantly enhancing the prediction accuracy. In terms of the practicality of the photovoltaic installation potential assessment framework, the constructed photovoltaic installation potential assessment framework performs excellently in practical applications. This framework comprehensively considers various factors such as the available roof area, sunshine conditions, building types, and geographical locations, and can accurately evaluate the installation potential of rooftop distributed photovoltaic systems. The experimental results show that the evaluation results provided by this framework are scientific and accurate, with high practicality and promotion value. In terms of supporting the urban energy structure transformation, the experimental results show that through accurate electricity demand prediction and reasonable photovoltaic installation potential assessment, it is possible to effectively guide the layout and scale optimization of distributed photovoltaic power generation systems, ensuring that the system can not only meet the electricity demand of users but also achieve efficient utilization of resources. This provides strong support for the transformation of the urban energy structure towards a clean and low-carbon direction, helps to improve the urban energy self-sufficiency rate, reduce carbon emissions, and improve environmental quality. In terms of future research directions, although the experimental results are encouraging, there are still some issues that need further research. For example, the generalization ability of the model under different climate conditions needs to be further verified, more external factors (such as electricity price policies, user behavior habits, etc.) should be incorporated into the evaluation framework, and how to achieve dynamic monitoring and intelligent management of distributed photovoltaic systems. Future research will focus on these issues to further improve the performance and benefits of distributed photovoltaic systems.

[0105] This application uses publicly available datasets for experiments. The experimental results show that the proposed composite prediction model and evaluation framework still have high accuracy and reliability when dealing with data from different regions. This provides strong technical support and theoretical basis for future applications in Shiyan City and other similar cities.

[0106] The present invention relates to a large-scale rooftop distributed photovoltaic (PV) installation potential assessment technology based on deep learning, which is particularly suitable for applications in urban environments. It solves the problem of insufficient prediction accuracy of user electricity demand in the prior art and discloses a composite prediction model combining Variational Mode Decomposition (VMD) and Crossformer network. This model first uses VMD to preprocess the original electricity data, effectively separating modal components of different frequencies to reduce noise interference and improve data quality; then it is further analyzed through the Crossformer network to capture the long-term dependencies in the time series, thereby achieving high-precision prediction of electricity demand.

[0107] In addition, the present invention also designs a comprehensive evaluation framework for the potential of photovoltaic installations, which takes into account factors including but not limited to the available roof area, sunshine conditions, building types, and geographical locations, aiming to accurately estimate the potential installation capacity of rooftop distributed photovoltaic systems. This method not only improves the accuracy and practicality of the evaluation, but also, due to its flexibility and adaptability, can be widely applied to different urban environments, such as Shiyan City, etc., providing a scientific basis and technical support for urban energy planning.

[0108] The evaluation method and model proposed by the present invention are of great significance for promoting the application and development of clean energy, especially for promoting the wide application of distributed photovoltaic power generation systems in cities.

[0109] The above research on the method for evaluating the potential of large-scale rooftop distributed photovoltaic installations in Shiyan based on deep learning provided by the present invention has been elaborated in detail. Specific examples are used in this application to illustrate the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those skilled in the art, there will be changes in the specific implementation manner and application scope according to the idea of the present invention. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for evaluating the installation potential of rooftop distributed photovoltaic based on deep learning, characterized in that, Including: Obtain a publicly available electricity demand dataset, decompose the original electricity demand data based on the VMD algorithm to obtain multiple sub-modal data; Based on the original electricity demand data and the multiple sub-modal data, obtain reconstructed data, and perform data partitioning on the reconstructed data to obtain a training set, a validation set, and a test set; Obtain a Crossformer network model, input the training set and the validation set into the Crossformer network model for iterative training to obtain a deep learning model for electricity demand prediction; Input the test set into the deep learning model to obtain an electricity demand prediction result.

2. The method for evaluating the installation potential of rooftop distributed photovoltaic based on deep learning according to claim 1, wherein The decomposing the original electricity demand data based on the VMD algorithm to obtain multiple sub-modal data includes: Obtain the decomposition layer number, and sequentially decompose the original electricity demand data according to the decomposition layer number based on the VMD algorithm to obtain multiple sub-modal data.

3. The method for evaluating the installation potential of rooftop distributed photovoltaics based on deep learning according to claim 2, characterized in that The obtaining the decomposition layer number includes: Adaptive adjustment of the decomposition layer number based on the correlation coefficient to obtain the optimal decomposition layer number.

4. A method for evaluating the installation potential of rooftop distributed photovoltaic based on deep learning according to claim 3, characterized in that After performing VMD decomposition on the original data, K modal components are obtained. Calculate the correlation coefficient between each modal component and the original signal to measure the corresponding correlation degree; when the absolute value of the correlation coefficient is less than 0.2, it is considered that there is almost no correlation between the two variables; when the absolute value of the correlation coefficient is higher than 0.2, it indicates that each modal component is an effective component and the signal is still in an under-decomposed state, and the value of K needs to be increased; when the absolute value of the minimum correlation coefficient first drops below 0.2, it is considered that the important information in the signal has been fully decomposed, and the K at this time is the optimal decomposition layer number.

5. A method for evaluating the installation potential of rooftop distributed photovoltaics based on deep learning according to claim 1, characterized in that, The electricity demand prediction model includes: Dimensional segment embedding, two-stage attention layer, and feature fusion module; Input the test set into the electricity demand prediction model, segment the test set through the dimensional segment embedding to obtain multiple segmented data, obtain feature vectors, and embed the multiple segmented data into the feature vectors to obtain multiple two-dimensional vector arrays. Capture features of the two-dimensional vector arrays through the two-stage attention layer to obtain captured features, and merge the multiple two-dimensional vector arrays based on the captured features through the feature fusion module to obtain multiple merged vectors, and obtain an electricity demand prediction result based on the multiple merged vectors.

6. The method for evaluating the installation potential of rooftop distributed photovoltaics based on deep learning according to claim 5, wherein, The feature fusion module includes: a three-layer encoder-decoder unit; Merge and predict the multiple two-dimensional vector arrays based on the captured features through the three-layer encoder-decoder unit in the feature fusion module to obtain multiple merged vectors.

7. A method for evaluating the installation potential of rooftop distributed photovoltaics based on deep learning according to claim 6, characterized in that, The three-layer encoder-decoder unit includes: a first-layer encoder-decoder unit; The first-layer encoder-decoder unit includes: an encoder and a decoder; the output end of the first-layer encoder is connected to the input end of the first-layer decoder; The first encoder in the first encoder-decoder unit of the three-layer encoder-decoder unit merges the multiple two-dimensional vector data based on the captured features to obtain a feature representation of the first encoder. The first decoder in the first encoder-decoder unit of the three-layer encoder-decoder unit predicts the feature representation of the first encoder to obtain a first merged vector.

8. A method for evaluating the installation potential of rooftop distributed photovoltaic based on deep learning according to claim 5, characterized in that Obtaining an electricity demand prediction result based on the multiple merged vectors includes: Performing an addition process on the multiple merged vectors to obtain an electricity demand prediction result.

9. A method for evaluating the installation potential of rooftop distributed photovoltaic based on deep learning, characterized in that, It includes: A sub-modal data unit for obtaining the original electricity demand data, decomposing the original electricity demand data based on the VMD algorithm to obtain multiple sub-modal data; A data set division unit for obtaining reconstructed data based on the original electricity demand data and the multiple sub-modal data, and performing data division on the reconstructed data to obtain a training set, a validation set, and a test set; A unit for constructing an electricity demand prediction model for obtaining a Crossformer network model, inputting the training set and the validation set into the Crossformer network model for iterative training to obtain an electricity demand prediction model; An output unit for inputting the test set into the electricity demand prediction model to obtain an electricity demand prediction result.

Images