A photovoltaic project survey data integration and analysis platform

Abstract

The application relates to a photovoltaic project survey data integration and analysis platform, which learns the relevance between semantic analysis, image features and data scores of survey data in different regions by adopting an artificial intelligence model, and serves as a correlation mode of survey data in different regions; a deep learning model between survey data and photovoltaic power generation is trained by using historical data, and the deep learning model is adaptively adjusted according to the correlation mode of survey data of a photovoltaic project to be evaluated; the application solves the problem that the prior art cannot consider the difference of the correlation mode between survey data in different regions, and the problem that the model constructed is inaccurate in evaluating the power generation potential of a photovoltaic project.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power generation technology and relates to a photovoltaic project survey data integration and analysis platform. Background Technology

[0002] With the increasing global demand for clean energy, photovoltaic (PV) projects, as an important component of the renewable energy sector, are experiencing rapid development. In the process of developing PV projects, the analysis of preliminary survey data can assess and predict the potential of PV power generation, playing a decisive role in the planning, construction, and operation of the project.

[0003] However, current technologies have several shortcomings. Firstly, data integration is poor. Preliminary survey data comes from a wide range of sources, including meteorology, geographic information, and field measurements. Existing technologies struggle to deeply correlate and analyze data from these different sources (regions), resulting in fragmented and isolated data that cannot provide comprehensive support for project decision-making. Secondly, analytical models lack adaptability. Photovoltaic projects operate in complex and variable environments, with significant differences in climate, topography, and geology across different regions. Existing analytical models are mostly built based on fixed parameters, making it difficult to automatically adjust to the unique environmental characteristics of each project. This leads to poor accuracy and reliability of the analysis results, hindering precise guidance for each stage of the project. Summary of the Invention

[0004] This invention provides a photovoltaic project survey data integration and analysis platform, which aims to achieve in-depth data fusion analysis and adaptive model adjustment, and solve the problem of inaccurate prediction of photovoltaic power generation potential caused by differences in the correlation patterns of survey data from different photovoltaic projects.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] This application provides a photovoltaic project survey data integration and analysis platform, including a multi-source data acquisition module, a multi-source data fusion module, an adaptive model building module, and a data analysis and prediction module. The multi-source data acquisition module, multi-source data fusion module, adaptive model building module, and data analysis and prediction module are communicatively connected, wherein:

[0007] The multi-source data acquisition module is used to collect survey data of the area where the photovoltaic project is located;

[0008] The multi-source data fusion module is used to analyze and process survey data using deep fusion technology and identify the correlation patterns of the survey data.

[0009] The adaptive model building module is used to train a deep learning model between survey data and photovoltaic power generation using historical data, and to adaptively adjust the deep learning model according to the correlation pattern.

[0010] The data analysis and prediction module is used to input the survey data of the photovoltaic project to be evaluated into the deep learning model and output the predicted photovoltaic power generation.

[0011] Furthermore, the survey data includes meteorological data, geographical data, equipment data, and aerial images.

[0012] Furthermore, the deep fusion technology includes:

[0013] S1. Obtain a sample dataset of photovoltaic project survey data from different regions during historical periods;

[0014] S2. Use NLP technology to perform semantic analysis on historical survey data;

[0015] S3. Use a convolutional neural network model to extract image features from historical survey data;

[0016] S4. Principal component analysis is used to calculate the data score of historical survey data;

[0017] S5. An artificial intelligence model is used to learn the correlation between semantic parsing, image features and data scores of historical survey data from different regions, which serves as the correlation pattern after the fusion of survey data from different regions.

[0018] Furthermore, the semantic parsing of historical survey data using NLP technology includes the following steps:

[0019] Data preprocessing: Text cleaning, normalization, and word segmentation are performed on historical survey data;

[0020] Part-of-speech tagging: Each processed word segment is labeled with a part-of-speech tag to clarify the grammatical relationships between words in the sentence;

[0021] Named entity recognition: Extracting named entities related to photovoltaic project surveying from word segmentation;

[0022] Syntactic analysis: By analyzing the grammatical structure of a sentence through dependency parsing, and analyzing the dependency relationships between word parts, the semantics of the sentence can be understood.

[0023] Semantic role labeling: Based on syntactic analysis, we delve into the semantic relationships between predicates and other components in a sentence to clarify the semantic roles of each component;

[0024] Knowledge fusion and semantic understanding: Integrate the information obtained from the previous steps and deeply integrate it with the professional knowledge map of the photovoltaic field to transform text data into structured knowledge related to the survey and analysis of photovoltaic projects.

[0025] Furthermore, the step of extracting image features from historical survey data using a convolutional neural network model includes the following steps:

[0026] Data preparation: Resize the image data of historical survey data to standard size, convert it to color or grayscale, and then normalize the pixel values;

[0027] Constructing a convolutional neural network model: Based on the task and data characteristics, define network layers, specifically including convolutional layers for extracting local features, pooling layers for downsampling to reduce data volume, fully connected layers for mapping features, and activation functions to enhance nonlinear expressions;

[0028] Model training: The preprocessed image data is divided into training set, validation set and test set according to the proportion. The learning rate, batch size and number of training rounds are set to train the parameters. The model is trained with the training set. The loss is calculated through forward propagation and the parameters are updated through back propagation. The performance of the model is evaluated and the parameters are tuned with the validation set during training. Finally, the generalization ability of the model is evaluated with the test set.

[0029] Feature extraction: After the model training reaches the target, historical survey images are input into the model, and the selected feature layer is output through forward propagation, which is the image feature.

[0030] Furthermore, the calculation of data scores for historical survey data using principal component analysis includes the following steps:

[0031] Data standardization: The Z-score standardization method is used to standardize the values ​​of each variable;

[0032] Calculate the covariance matrix: After data standardization, calculate the covariance matrix for historical survey data;

[0033] Solving for eigenvalues ​​and eigenvectors: Expanding the covariance matrix to find the eigenvalues ​​and eigenvectors;

[0034] Principal component selection: Principal components are selected based on the solved eigenvalues ​​and eigenvectors;

[0035] Calculate the data score: Calculate the data score using the selected principal components and standardized historical survey data.

[0036] Furthermore, the method of using an artificial intelligence model to learn the correlation between semantic parsing, image features, and data scores of historical survey data from different regions includes the following steps:

[0037] Data integration and preparation: The semantic analysis results, image features, and data scores of historical survey data from different regions are comprehensively integrated and categorized by region;

[0038] Model selection: Based on the characteristics and complexities of the survey data from different regions, select an artificial intelligence model;

[0039] Model training: Using semantic parsing, image features, and data scores from different regions as input, the model learns the correlation patterns between data from different regions.

[0040] Furthermore, the deep learning model is configured as a multilayer perceptron, and includes the following construction steps:

[0041] Determine the input layer: Determine the number of neurons in the input layer based on the dimensions of the survey data, and use the raw data as input to provide basic information for model learning;

[0042] Designing hidden layers: The number of hidden layers and the number of neurons in each layer are determined through experiments and experience. Layers are connected by weight matrices, and neurons are subjected to weighted summation and nonlinear transformation by activation functions.

[0043] Constructing the output layer: The output layer has one neuron that receives the output of the last hidden layer. The predicted value is obtained by weighted summation and a linear activation function.

[0044] Initialize weights and biases: After the model is built, the weight matrices of each layer and the biases of the neurons are initialized. The initialization is done by randomly initializing the weights.

[0045] Choosing the loss function and optimizer: The mean squared error loss function is selected to measure the difference between the predicted and actual power generation. Stochastic gradient descent is used to adjust the weight bias to minimize the loss value.

[0046] Furthermore, the adaptive adjustment of the deep learning model based on the association pattern includes the following steps:

[0047] Correlation pattern analysis: Using visualization and statistical analysis methods, we can clarify the interactions between survey data and their impact on power generation.

[0048] Feature importance assessment: Based on the association pattern, a random forest model is used to determine the importance of different combinations of survey data features in predicting power generation, and to quantify the influence of each survey data feature.

[0049] Model structure adjustment: Based on the feature importance evaluation results, the deep learning model structure is optimized, specifically by adding neurons or nodes for important features and reducing the relevant parts of unimportant features;

[0050] Weight and parameter adjustment: Increase the adjustment range of the weights and parameters of important features;

[0051] Model validation and iteration: Use validation datasets to test the adjusted model and evaluate the model's predictive performance using evaluation metrics.

[0052] The beneficial effects of this invention are:

[0053] (1) An artificial intelligence model is used to learn the correlation between semantic parsing, image features and data scores of survey data in different regions as the correlation pattern of survey data in different regions; a deep learning model between survey data and photovoltaic power generation is trained using historical data, and the deep learning model is adaptively adjusted according to the correlation pattern of the survey data of the photovoltaic project to be evaluated; the present invention solves the problem that the existing technology cannot consider the differences in the correlation patterns between survey data in different regions, which leads to the inaccurate evaluation of the power generation potential of photovoltaic projects by the constructed model.

[0054] (2) The survey data includes meteorological data, geographical data, equipment data and aerial images. The identified survey data association patterns can accurately characterize the differences in power generation in different regions, thus improving the accuracy of photovoltaic project evaluation.

[0055] (3) By utilizing the influence of different survey data characteristics on photovoltaic power generation in different regions, the model is adaptively adjusted to improve the accuracy of model prediction. Attached Figure Description

[0056] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0057] Figure 1 This is a structural diagram of a photovoltaic project survey data integration and analysis platform according to the present invention.

[0058] Figure 2 This is a flowchart illustrating the deep fusion technology analysis in one embodiment of the present invention.

[0059] Figure 3 This is a flowchart illustrating the adaptive adjustment of a deep learning model based on an association pattern in one embodiment of the present invention. Detailed Implementation

[0060] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0061] Please see Figures 1-3 This application provides a photovoltaic project survey data integration and analysis platform, including a multi-source data acquisition module, a multi-source data fusion module, an adaptive model building module, and a data analysis and prediction module. The multi-source data acquisition module, multi-source data fusion module, adaptive model building module, and data analysis and prediction module are communicatively connected, wherein:

[0062] The multi-source data acquisition module is used to collect survey data of the area where the photovoltaic project is located;

[0063] Furthermore, the survey data includes meteorological data, geographical data, equipment data, and aerial images.

[0064] In this embodiment, meteorological data includes information such as light intensity, sunshine duration, temperature, wind speed, and wind direction. Light intensity and sunshine duration directly determine the total amount of solar energy received and converted into electrical energy by the photovoltaic modules, and are key factors affecting power generation; temperature changes affect the photoelectric conversion efficiency of the modules, and excessively high or low temperatures may reduce power generation; wind speed and wind direction are related to the heat dissipation and stability of the equipment, and strong winds may put pressure on structures such as photovoltaic supports.

[0065] Geographic data includes topography, altitude, latitude and longitude. Topography affects the feasibility and layout of power plant construction. Flat terrain is conducive to large-scale installation, while complex terrain requires consideration of slope and aspect to optimize component arrangement. Altitude affects atmospheric transparency and light intensity. The higher the altitude, the greater the light intensity. Latitude and longitude determine the solar altitude angle and sunshine duration, which affect power generation.

[0066] Equipment data encompasses parameters and operational status information for various equipment in photovoltaic projects, such as the type, power, and conversion efficiency of photovoltaic modules; the conversion efficiency and maximum power point tracking accuracy of inverters; and operational status data like temperature, current, and voltage. This data directly determines the equipment's power generation capacity, energy conversion efficiency, and operational stability, and is crucial for ensuring the stable operation of the power generation system.

[0067] Aerial imagery provides a macroscopic view of photovoltaic projects, visually showcasing the overall site conditions, including topography, surrounding environment, and building distribution. By analyzing aerial images, shadowed areas can be quickly identified, such as the impact of surrounding buildings and trees on the power station, allowing for optimized planning and layout. It can also assist in assessing land use and determining suitable areas for module installation.

[0068] The multi-source data fusion module is used to analyze and process survey data using deep fusion technology and identify the correlation patterns of the survey data.

[0069] In this embodiment, deep fusion technology aims to comprehensively analyze photovoltaic project survey data and uncover the inherent relationships between different types of data. Given the differences in data correlation patterns among photovoltaic projects in different regions, this technology, through multi-step operations, accurately extracts correlation patterns applicable to specific projects, laying a solid foundation for subsequent data processing and analysis. This includes:

[0070] S1. Obtaining Historical Data Samples of Photovoltaic Project Surveys in Different Regions: This step is the starting point for deep integration technology and requires the extensive collection of photovoltaic project survey data from different historical periods and diverse geographical locations. These data sources are wide-ranging, covering meteorological data (such as sunlight intensity, temperature, and precipitation), geographical data (including altitude and topography), equipment data (such as photovoltaic module specifications and inverter performance parameters), and aerial imagery. Comprehensive and accurate sample data is the cornerstone of subsequent analysis. Sample data from different regions can reflect the characteristics of photovoltaic projects in various places, such as the differences between high-altitude and low-altitude areas, and between arid and humid areas. These differences are crucial for revealing the different correlation patterns in survey data from different locations.

[0071] S2. Using NLP technology for semantic analysis of historical survey data: Among numerous survey data, multi-source data may contain textual information. For weather forecast text in meteorological data, NLP technology uses word segmentation to break down phrases such as "tomorrow, localized areas will experience thunderstorms, with expected wind speeds of 5-7" into lexical units such as "tomorrow," "localized areas," "expected," "thunderstorms," ​​"expected," "wind speed," and "5-7." Part-of-speech tagging determines the part of speech of each word, such as "thunderstorms" being a noun and "5-7" being a quantifier. Named entity recognition identifies key information, such as the weather phenomenon "thunderstorms" and the wind speed value "5-7." Semantic role tagging clarifies the semantic role of words, such as "5-7" being the specific value of "wind speed." In geographic data-related text, such as descriptions of topography in project site selection reports, NLP technology can extract key information such as terrain type and topographic relief. In equipment data text, such as equipment maintenance manuals, NLP technology can parse out content such as equipment fault types and maintenance cycles.

[0072] Furthermore, the semantic parsing of historical survey data using NLP technology includes the following steps:

[0073] Data preprocessing: This is the first step in semantic parsing of historical survey data using NLP technology. Because historical survey data comes from a wide range of sources and takes many forms, the text often contains various kinds of interfering information. Text cleaning aims to remove noise such as garbled characters, special characters, and HTML tags from web page text, restoring the text to a clean state. Normalization focuses on unifying the format of key information such as numbers, dates, and times in the text, and restoring abbreviations to avoid comprehension barriers caused by format differences or abbreviated expressions. Word segmentation, as a crucial part of data preprocessing, uses specialized tools to accurately segment continuous text into independent words or units based on different language characteristics, enabling computers to process text with finer granularity and laying the foundation for subsequent steps such as part-of-speech tagging and entity recognition.

[0074] Part-of-speech tagging: After data preprocessing, part-of-speech tagging assigns a part-of-speech label to each word segment, clarifying its role in the grammatical system, such as noun, verb, adjective, etc. This process is crucial for a deeper understanding of text structure and semantic relationships. Through part-of-speech tagging, we can gain insight into the grammatical connections between words in a sentence, such as identifying the main components like subject, predicate, and object. Taking the photovoltaic project text "inverter continuously outputs stable current" as an example, "inverter" is tagged as a noun, serving as the subject of the action; "output" is tagged as a verb, acting as the predicate to represent the core action; and "current" is also a noun, representing the object of the action. Part-of-speech tagging tools are mostly based on statistical models or deep learning models, which can learn the part-of-speech rules of words based on a large amount of text data, thereby accurately tagging parts of speech.

[0075] Named Entity Recognition (NER): This step focuses on accurately extracting entities with specific meanings from text. These entities are closely related to photovoltaic project surveys and cover key information such as project location, equipment model, measurement time, and meteorological data. Deep learning-based methods, such as the Bi-LSTM + CRF model, perform excellently in handling complex text structures and diverse entity types, effectively recognizing various entities and providing crucial information support for subsequent semantic analysis.

[0076] Syntactic analysis aims to dissect the grammatical structure of a sentence and clarify the dependency relationships between its components, such as subject-verb-object, attributive, adverbial, and complement. Through this analysis, the sentence's organizational structure and semantic logic become clear, much like building a skeleton for the text. Dependency parsing and phrase structure analysis are commonly used methods. Dependency parsing clearly demonstrates the dependency relationships between words, aiding in understanding sentence semantics and laying a solid foundation for deeper semantic analysis.

[0077] Semantic role labeling: This step, based on syntactic analysis, delves into the semantic relationships between predicates (mostly verbs) and other components in a sentence, clarifying the semantic roles of each component, such as agent, patient, time, and place. Semantic role labeling relies on deep learning models and combines part-of-speech tagging, named entity recognition, and syntactic analysis results to accurately predict the semantic roles of each component, contributing to a comprehensive understanding of the text's semantic meaning.

[0078] Knowledge Fusion and Semantic Understanding: As the final step in semantic parsing, this stage integrates the information obtained from the previous steps and deeply fuses it with a professional knowledge graph in the photovoltaic field. For example, it associates the identified equipment model with detailed parameters, performance characteristics, and common faults of that model in the knowledge graph; it links the project location with geographical features, meteorological conditions, and policy environment of that region in the knowledge graph. Through this fusion, textual information is no longer isolated but interwoven with domain knowledge, thereby uncovering deeper semantics behind the text, such as potential causal relationships between different factors and best practice recommendations under specific conditions. This transforms textual data into structured knowledge with practical value for photovoltaic project surveying, analysis, and decision-making.

[0079] S3. Using Convolutional Neural Networks (CNNs) to extract image features from historical survey data: For aerial images, which directly present the conditions of a project area, CNNs play a crucial role. Through operations such as convolutional layers and pooling layers, they automatically extract key features from the images, such as terrain undulations, vegetation cover, and the distribution of surrounding buildings. Aerial images from different regions have unique characteristics due to varying geographical environments. CNNs can capture these differences; for example, aerial images of mountainous areas exhibit complex topographic features, while images of plains are relatively flat and open. Combining these features with other data helps to discover unique correlations between image features from different regions and other survey data, such as the relationship between mountainous terrain features and sunlight shading and power generation.

[0080] Furthermore, the step of extracting image features from historical survey data using a convolutional neural network model includes the following steps:

[0081] Data preparation: First, a wide range of historical survey images of photovoltaic projects, such as aerial photographs and equipment diagrams, are collected and organized according to rules. Then, the original images are preprocessed, resized to standard dimensions, converted to color or grayscale, and pixel values ​​are normalized to prepare for input to the convolutional neural network model, ensuring that the data retains key information while meeting the model's processing requirements.

[0082] Constructing a convolutional neural network model: Based on the task and data characteristics, select a suitable architecture, such as classic models like VGG and ResNet. On this basis, define network layers, including convolutional layers to extract local features, pooling layers to downsample and reduce data volume, and fully connected layers to map features. Activation functions are then used to enhance non-linear expression, building a network structure capable of effectively learning image features.

[0083] Model training: The preprocessed image data is divided into training, validation, and test sets according to a set ratio. Training parameters such as learning rate, batch size, and number of training epochs are set. The model is trained using the training set, and the loss is calculated through forward propagation. The parameters are updated through backpropagation. The performance is evaluated and the parameters are tuned using the validation set during training to prevent overfitting. Finally, the generalization ability of the model is evaluated using the test set.

[0084] Feature extraction: After the model training reaches the required standard, it is loaded. Appropriate layers are selected from each network layer of the model as needed. Shallow layers extract low-level features, and deeper layers extract high-level features. Historical survey images are input into the model, and the selected feature layer outputs, i.e., image feature representations, after forward propagation. These are used for subsequent photovoltaic project data analysis and mining of potential connections with other data.

[0085] S4. Principal Component Analysis (PCA) is used to calculate data scores for historical survey data: Survey data contains multiple variables, and there may be complex relationships between these variables. PCA aims to reduce the dimensionality of the data, remove redundant information, and extract key components. The original data is projected onto new orthogonal coordinate axes through linear transformation, and principal components are determined based on the variance. Data scores for each sample on the principal components are calculated; these scores summarize the core characteristics of the sample. The relationships between data variables differ across regions. For example, in some high-temperature areas, temperature is closely related to power generation, while in high-altitude areas, the correlation between altitude and sunlight intensity is more significant. PCA helps to clarify these regional differences in data relationships and highlight the main characteristic combinations of data from different regions.

[0086] Furthermore, the calculation of data scores for historical survey data using principal component analysis includes the following steps:

[0087] Data standardization: Historical survey data encompasses diverse variables with significant differences in dimensions and numerical ranges. Direct analysis could lead to some variables dominating the results. Therefore, methods such as Z-score standardization are employed to transform the values ​​of each variable into standard normal distribution data with a mean of 0 and a standard deviation of 1. This eliminates the influence of dimensions and ensures that all variables are on a uniform scale, laying a standardized data foundation for subsequent principal component analysis.

[0088] Calculate the covariance matrix: After data standardization, calculate the covariance matrix for historical survey data. This matrix displays the covariance values ​​between different variables through its elements, comprehensively presenting the linear correlation between various types of data such as meteorological, geographical, and equipment data. It helps to clarify the degree of influence of variable combinations on the overall characteristics of the data and provides core information for extracting principal components.

[0089] Finding eigenvalues ​​and eigenvectors: The covariance matrix is ​​expanded to find its eigenvalues ​​and eigenvectors. Eigenvalues ​​reflect the information content of each principal component, while eigenvectors determine the orientation of the principal components. In historical survey data, the principal components corresponding to larger eigenvalues ​​contain key information and may be linear combinations of multiple original variables. This step extracts the orientation and information content of key principal components from the internal structure of the data.

[0090] Principal component selection: Principal components are selected based on the solved eigenvalues ​​and eigenvectors. Generally, they are sorted by eigenvalue size, and the top few principal components with a cumulative contribution rate reaching a specific threshold (e.g., 80%-90%) are selected. The cumulative contribution rate represents the proportion of information contained in the selected principal components. In photovoltaic project survey data, appropriate principal component selection can retain key information, reduce dimensionality, and avoid excessive information loss, without affecting subsequent analysis.

[0091] Calculating Data Scores: Finally, the data scores are calculated using the selected principal components and standardized historical survey data. The standardized data are then multiplied by the principal component eigenvectors to obtain the scores for each sample on each principal component. These scores reflect the sample's position in the new principal component space, comprehensively demonstrating the sample's characteristics under the original variable combinations, and providing a quantitative analysis basis for each stage of the photovoltaic project.

[0092] S5. An artificial intelligence model is employed to learn the correlations between semantic parsing, image features, and data scores of historical survey data from different regions, serving as the correlation pattern after data fusion. Using AI models (such as deep neural networks and random forests), the model takes semantically parsed structured data, image features extracted by CNNs, and data scores calculated by PCA as input. Through extensive training, the model learns the mapping relationships between these different types of features. Since the input data features differ across regions, the model learns unique correlation patterns for each location. In the in-depth analysis of historical survey data for photovoltaic projects, a multi-dimensional and close correlation exists between semantic parsing, image features, and data scores.

[0093] From a meteorological and geographical perspective, semantically parsed meteorological text data, such as descriptions of sunshine duration, intensity, and precipitation frequency and magnitude, shows a significant correlation with geographical features obtained through aerial imagery. For example, if semantic parsing indicates that a region has "abundant sunshine year-round and heavy snowfall in winter," and the aerial imagery shows the area as a flat plain with open terrain and no tall obstructions, this explains the abundant sunshine. Conversely, if the imagery shows surrounding mountains and the project being located on the leeward side, it might correspond to the description of heavy snowfall in winter. Principal component analysis (PCA) generates a data score that comprehensively considers the climatic conditions conveyed by the meteorological text semantics and the geographical features reflected in the imagery, thus generating a score reflecting the overall meteorological and geographical suitability of the region for photovoltaic (PV) projects. A higher score indicates that the region's meteorological and geographical conditions are favorable for PV projects; for example, abundant sunshine and suitable terrain facilitate efficient power generation by PV modules. Conversely, a lower score may suggest problems affecting the project, such as heavy snowfall in winter causing snow accumulation and shading of PV modules, impacting power generation efficiency.

[0094] Regarding equipment performance and environmental factors, semantic analysis of texts such as equipment maintenance manuals and fault reports reveals information such as the type, frequency, and cause of equipment faults (e.g., "the output power of photovoltaic modules decreased due to prolonged high temperatures"), which is closely linked to the surrounding environmental conditions presented in image features. For example, thermal imaging images may show excessively high surface temperatures of photovoltaic modules, echoing the text's mention of high temperatures causing faults. Data scoring integrates semantic information related to equipment performance from the text and environmental features reflected in the images. For instance, it combines equipment fault conditions with factors such as ambient temperature, humidity, and wind and sand to arrive at a data score reflecting the equipment's operational stability in the current environment. A lower data score indicates that the equipment may frequently fail in the current environment, requiring further optimization of the equipment's heat dissipation system, strengthened protective measures, or replacement with a more environmentally suitable equipment model to ensure the stable operation of the photovoltaic project.

[0095] In a time series analysis, the semantically parsed text data includes information such as weather forecasts, equipment status updates, and project progress records at different points in time. Image features also exhibit dynamic changes over time, such as the impact of vegetation growth and withering in different seasons on the shading of photovoltaic modules, and changes in the appearance and aging of equipment during long-term use. Data scores also change due to time factors. In-depth analysis of the correlations between these data over time can uncover long-term development trends and inherent patterns. For example, over time, semantic parsing of equipment maintenance records can reveal signs of equipment aging, such as a gradual increase in failure frequency and maintenance complexity. Simultaneously, the appearance of the equipment in the images may show obvious signs of aging, such as wear and discoloration. Data scores will also gradually decrease accordingly, comprehensively reflecting the deterioration of the overall operating condition of the equipment. This time series correlation analysis is of great significance for predicting the future development trend of photovoltaic projects. By learning the correlation patterns of historical data over time, artificial intelligence models can accurately predict future changes in equipment performance, fluctuations in power generation, and the probability of failures. This provides a scientific basis for project operation and maintenance decisions, such as developing equipment replacement plans in advance and optimizing maintenance cycles. It also provides strong support for resource allocation, such as rationally arranging human and material resources to ensure that photovoltaic projects always maintain a high-efficiency and stable operating state.

[0096] However, this correlation varies across different regions. For example, compared to low-altitude areas, high-altitude areas show a stronger positive correlation between sunlight and power generation, and temperature has a greater weight in influencing equipment performance. Therefore, this embodiment uses an artificial intelligence model to learn the correlation between semantic parsing, image features, and data scores of survey data from different regions to determine the correlation between survey data from different regions.

[0097] In the entire photovoltaic project survey data processing workflow, the preceding steps performed multi-dimensional analysis of historical survey data, including sample acquisition, NLP semantic parsing, convolutional neural network image feature extraction, and principal component analysis to calculate data scores. Step S5 focuses on using artificial intelligence models to specifically learn the correlations between the aforementioned data types in different regions, thereby obtaining survey data correlation patterns applicable to different regions. This makes the analysis results more region-specific. This includes:

[0098] Data Integration and Preparation: This involves comprehensively integrating the semantic analysis results, image features, and data scores of historical survey data obtained from previous steps across different regions. Each region's data includes information extracted from the text, key features extracted from the images, and data scores derived from comprehensive analysis. For example, meteorological text from region A, after semantic analysis, reveals unique climate characteristics; its images present specific geographical features; and it also includes data scores based on various factors. Region B similarly has different data. This data is then categorized and organized by region to form a dataset suitable for processing by artificial intelligence models.

[0099] Choosing the right AI model: Given the complexity of data characteristics and relationships across different regions, it's crucial to select an AI model capable of effectively capturing these differences. For situations with complex and non-linear data relationships, deep neural network models such as Multilayer Perceptrons (MLPs), Recurrent Neural Networks (RNNs) and their variants, Long Short-Term Memory Networks (LSTMs) and Gated Recurrent Units (GRUs) may be more suitable. For example, if there are complex dynamic relationships between meteorological, geographical, and equipment data from different regions, LSTMs or GRUs can better handle such time-series or sequential data relationships. If the relationships between data are relatively simple, traditional machine learning models such as decision trees and random forests can also be effective; they can identify patterns in the relationships between data from different regions through the analysis of data features.

[0100] Model Training: The selected AI model is trained using integrated datasets from different regions. During training, the model takes semantic parsing, image features, and data scores from different regions as input, learning the unique correlation patterns between data from each region. The model continuously adjusts its parameters to minimize the error between the predicted correlation patterns and the actual correlation patterns. For example, for a specific region, the model predicts the correlation between light intensity, terrain features, and power generation, compares this prediction with the actual correlation patterns in the data, and updates the parameters using a backpropagation algorithm to make the predictions closer to reality. The training process is repeated on data from different regions, allowing the model to gradually master the correlation patterns between data from different regions.

[0101] Association Pattern Learning and Output: As training progresses, the AI ​​model gradually learns the correlations between various features of historical survey data from different regions. These learned correlations constitute the association patterns of survey data from different regions. For example, the model might discover that the correlation between sunlight intensity and power generation is stronger in high-altitude areas than in low-altitude areas, or that the impact of humidity on equipment performance differs between coastal and inland areas. These association patterns exist in the form of model parameters and internal representations, and can be used to analyze and predict new survey data from the same region, providing targeted decision-making support for the planning, construction, and operation and maintenance of photovoltaic projects in different regions.

[0102] The adaptive model building module is used to train a deep learning model between survey data and photovoltaic power generation using historical data, and to adaptively adjust the deep learning model according to the correlation pattern.

[0103] In this embodiment, the adaptive model building module aims to accurately construct and dynamically optimize the relationship model between survey data and photovoltaic power generation, focusing on training a deep learning model using historical data. A wide range of historical data related to photovoltaic projects are collected, including meteorological survey data (such as light intensity, temperature, and humidity); geographical survey data (such as altitude and topography); and equipment operation survey data (such as component power and inverter efficiency), along with corresponding photovoltaic power generation data. After collection, these data undergo preprocessing to remove outliers and noise, and standardization is used to unify the scale of data with different dimensions, laying a solid foundation for subsequent model training. Based on data characteristics and project requirements, suitable deep learning model architectures are carefully selected, such as multilayer perceptrons, recurrent neural networks, and their variants, and the model is built, determining key parameters such as the number of layers and neurons. Subsequently, the preprocessed data is used to train the model, with survey data as input and power generation as the output label. Optimization algorithms are used to continuously adjust the model parameters, gradually reducing the error between predicted and actual power generation, thereby constructing a preliminary prediction model.

[0104] Secondly, this module adaptively adjusts the deep learning model based on correlation patterns. These correlation patterns originate from multi-dimensional analysis of historical survey data, reflecting the inherent connections between various features of survey data from different regions. By deeply analyzing these correlation patterns and understanding the data relationships in different regions and their impact on power generation, the adaptive model building module makes targeted adjustments to the trained model. If it is found that topographic factors in a certain region have a significant impact on power generation, and the initial model does not adequately consider this, an input layer or hidden layer related to topographic features will be added; or the feature weights in the model will be adjusted according to the importance of each factor in the correlation pattern, or even the model structure will be adjusted according to the complexity of the correlation pattern. In this way, the model can more accurately fit the actual conditions of different regions, significantly improving the prediction accuracy and adaptability of photovoltaic power generation, and providing strong support for the efficient planning, operation and maintenance, and decision-making of photovoltaic projects.

[0105] Furthermore, the deep learning model is configured as a multilayer perceptron, and includes the following construction steps:

[0106] Determining the Input Layer: The input layer of a Multilayer Perceptron (MLP) must be built closely around the survey data of the photovoltaic project. This includes collecting meteorological data (such as light intensity, temperature, humidity, and wind speed), geographical data (such as altitude, slope, latitude, and longitude), and equipment-related data (such as the model and power of photovoltaic modules, and inverter efficiency). This data comprehensively reflects various factors affecting photovoltaic power generation. The number of neurons in the input layer is determined based on the dimensions of the data. For example, if a total of 10 different types of survey data are collected, then the input layer would have 10 neurons, each corresponding to a data feature. These raw data are used as input to the model, providing foundational information for subsequent learning and prediction.

[0107] Designing Hidden Layers: Hidden layers are the core of an MLP, responsible for learning complex patterns in the data. Determining the number of hidden layers and the number of neurons in each layer is crucial. This is generally chosen through experimentation and experience, typically starting with 1-3 hidden layers. The number of neurons per layer can be adjusted based on the number of neurons in the input layer and the complexity of the data. For example, if the data is complex, the first hidden layer might have more neurons than the input layer, such as 15-20, to increase the model's learning capacity. The number of neurons in subsequent hidden layers can be gradually reduced, such as 10-15 in the second layer and 5-10 in the third. Hidden layers are connected by weight matrices. Each neuron receives the output of the previous layer through a weighted summation and undergoes a non-linear transformation using an activation function (such as ReLU), enabling the model to learn non-linear relationships in the data and uncover the complex correlation between survey data and power generation.

[0108] Constructing the output layer: The construction of the output layer is relatively simple. Since the goal is to predict photovoltaic power generation, only one neuron is needed. This neuron receives the output of the last hidden layer and obtains the final prediction result through weighted summation and an activation function (a linear activation function, because power generation is a continuous value). This predicted value is the photovoltaic power generation predicted by the model based on the input survey data. The weights between the output layer and the last hidden layer are also optimized through training to make the predicted power generation as close as possible to the actual power generation, thereby achieving the goal of accurately predicting photovoltaic power generation using the MLP model.

[0109] Initializing Weights and Bias: After the model is built, the weight matrices between layers and the biases of neurons need to be initialized. The method of weight initialization affects the training effect and convergence speed of the model. Common initialization methods include random initialization, such as using a Gaussian or uniform distribution to randomly generate weight values, ensuring that the weights take values ​​within a certain range and avoiding excessively large or small values ​​that could lead to training difficulties. Bias is usually initialized to a small constant, such as 0 or a value close to 0. Reasonable initialization of weights and biases allows the model to start learning in a better state in the early stages of training, laying the foundation for subsequent optimization of weights and biases through the backpropagation algorithm, thereby improving the model's prediction accuracy.

[0110] Choosing a loss function and optimizer: The loss function measures the difference between the model's predicted and actual values. For MLP models predicting photovoltaic power generation, the mean squared error (MSE) loss function is a commonly used choice. It intuitively reflects the accuracy of the model's prediction by calculating the average of the squares of the difference between the predicted and actual power generation. The optimizer is responsible for adjusting the model's weights and biases based on the feedback from the loss function to minimize its value. Stochastic gradient descent (SGD) and its variants (such as Adagrad, Adadelta, Adam, etc.) are common optimizers. For example, the Adam optimizer combines the advantages of Adagrad and Adadelta, adaptively adjusting the learning rate, performing well during training, and enabling the model to converge to the optimal solution faster, thereby improving the model's predictive performance.

[0111] Furthermore, the adaptive adjustment of the deep learning model based on the association pattern includes the following steps:

[0112] Association Pattern Analysis: This involves in-depth research into the association patterns between semantic parsing, image features, and data scores in historical survey data from different regions. For example, it analyzes which factors significantly impact power generation in specific areas, such as the close relationship between sunlight intensity and power generation in high-altitude regions, or the influence of humidity and equipment performance on power generation in coastal areas. Through visualization tools (such as creating heat maps to show the correlation between factors) and statistical analysis (calculating correlation coefficients), the interactions between factors and their degree of influence on power generation are clearly presented, providing a clear basis for model adjustments.

[0113] Feature Importance Assessment: Based on association patterns, the importance of different features in predicting power generation is determined. Features that frequently appear in association patterns and have a significant impact on power generation are assigned higher importance. For example, in areas with abundant solar resources, solar intensity features are highly important; while in areas with complex terrain, terrain-related features are more prominent. Feature selection algorithms (such as recursive feature elimination) or machine learning-based feature importance assessment methods (such as feature importance ranking in random forests) are used to quantify the importance of each feature, allowing for targeted model adjustments later. In photovoltaic project data analysis, association patterns reflect the inherent connections between semantic parsing, image features, and data scores in historical survey data. Random forests are ensemble learning models composed of multiple decision trees. Using random forest models to determine the importance of different combinations of associated survey data features for predicting power generation based on association patterns is beneficial because it can handle high-dimensional data, automatically consider the interactions between features, and has good stability and generalization ability. By training random forest models, the influence of each survey data feature on power generation prediction can be quantified, thereby identifying key features and their associated combinations that have a significant impact on power generation.

[0114] Model Structure Adjustment: Based on the feature importance assessment results, the structure of the Multilayer Perceptron (MLP) is adjusted. If a key feature is not fully represented in the original model, related input layer neurons or hidden layer nodes can be added. For example, if vegetation cover in a certain area is found to have a significant impact on power generation, but the original model does not consider this feature, vegetation cover-related neurons can be added to the input layer, and nodes can be appropriately added to the hidden layer to enhance the learning ability for this feature. Conversely, neurons or connections corresponding to features with low importance can be reduced or deleted to simplify the model structure and avoid overfitting.

[0115] Weight and Parameter Adjustment: In addition to structural adjustments, the model's weights and parameters also need to be optimized based on the association patterns. For weights corresponding to highly important features, a larger adjustment range should be given during training to more accurately reflect the relationship between the feature and power generation. For example, by adjusting the learning rate, a larger learning rate can be applied to the weights related to important features to accelerate their convergence. Simultaneously, the model's hyperparameters (such as the number of hidden layers and neuron activation functions) should be re-examined and appropriately adjusted based on the characteristics of the association patterns and the model's performance on the validation set to improve the model's adaptability to data from different regions and its prediction accuracy.

[0116] Model Validation and Iteration: After completing the above adjustments, validate the model using a validation dataset. Run the adjusted model on the validation set and compare the predicted power generation with the actual power generation. Evaluate the model performance by calculating metrics such as mean squared error (MSE) and mean absolute error (MAE). If the model performance does not meet expectations, repeat the above steps to further analyze correlation patterns, adjust the model structure and parameters, and continuously iterate and optimize until the model demonstrates good adaptability and prediction accuracy on the validation set, accurately reflecting the relationship between survey data and power generation in different regions.

[0117] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A photovoltaic project survey data integration and analysis system, characterized by: It includes a multi-source data acquisition module, a multi-source data fusion module, an adaptive model building module, and a data analysis and prediction module. These modules are communicatively connected. The multi-source data acquisition module is used to collect survey data of the area where the photovoltaic project is located; the survey data includes meteorological data, geographical data, equipment data, and aerial images; The multi-source data fusion module is used to analyze and process survey data using deep fusion technology to identify correlation patterns in the survey data; the deep fusion technology includes: S1. Obtain a sample dataset of photovoltaic project survey data from different regions during historical periods; S2. Use NLP technology to perform semantic analysis on historical survey data; S3. Use a convolutional neural network model to extract image features from historical survey data; S4. Principal component analysis is used to calculate the data score of historical survey data; S5. Use an artificial intelligence model to learn the correlation between semantic parsing, image features and data scores of historical survey data from different regions, as the correlation pattern after the fusion of survey data from different regions; The adaptive model building module is used to train a deep learning model between survey data and photovoltaic power generation using historical data, and to adaptively adjust the deep learning model according to the correlation pattern, including the following steps: Correlation pattern analysis: Using visualization and statistical analysis methods, we can clarify the interactions between survey data and their impact on power generation. Feature importance assessment: Based on the association pattern, a random forest model is used to determine the importance of different combinations of survey data features in predicting power generation, and to quantify the influence of each survey data feature. Model structure adjustment: Based on the feature importance evaluation results, the deep learning model structure is optimized, specifically by adding neurons or nodes for important features and reducing neurons or nodes for unimportant features; Weight and parameter adjustment: Increase the adjustment range of the weights and parameters of important features; Model validation and iteration: Use the validation dataset to test the adjusted model and evaluate the model's predictive performance using evaluation metrics; The data analysis and prediction module is used to input the survey data of the photovoltaic project to be evaluated into the deep learning model and output the predicted photovoltaic power generation.

2. The photovoltaic project survey data integration and analysis system of claim 1, wherein: The semantic parsing of historical survey data using NLP technology includes the following steps: Data preprocessing: Text cleaning, normalization, and word segmentation are performed on historical survey data; Part-of-speech tagging: Each processed word segment is labeled with a part-of-speech tag to clarify the grammatical relationships between words in the sentence; Named entity recognition: Extracting named entities related to photovoltaic project surveying from word segmentation; Syntactic analysis: By analyzing the grammatical structure of a sentence through dependency parsing, and analyzing the dependency relationships between word parts, the semantics of the sentence can be understood. Semantic role labeling: Based on syntactic analysis, we delve into the semantic relationships between predicates and other components in a sentence to clarify the semantic roles of each component; Knowledge fusion and semantic understanding: Integrate the information obtained from the previous steps and deeply integrate it with the professional knowledge map of the photovoltaic field to transform text data into structured knowledge related to the survey and analysis of photovoltaic projects.

3. The photovoltaic project survey data integration and analysis system of claim 1, wherein: The method of extracting image features from historical survey data using a convolutional neural network model includes the following steps: Data preparation: Resize the image data of historical survey data to the standard size, convert color images to grayscale, and then normalize the pixel values; Constructing a convolutional neural network model: Based on the task and data characteristics, define network layers, specifically including convolutional layers for extracting local features, pooling layers for downsampling to reduce data volume, fully connected layers for mapping features, and activation functions to enhance nonlinear expressions; Model training: The preprocessed image data is divided into training set, validation set and test set according to the proportion. The learning rate, batch size and number of training rounds are set to train the parameters. The model is trained with the training set. The loss is calculated through forward propagation and the parameters are updated through back propagation. The performance is evaluated and the parameters are tuned with the validation set during training. Finally, the generalization ability of the model is evaluated with the test set. Feature extraction: After the model training reaches the target, historical survey images are input into the model, and the feature layer output, i.e., image features, is obtained through forward propagation.

4. The photovoltaic project survey data integration and analysis system of claim 1, wherein: The calculation of data scores for historical survey data using principal component analysis includes the following steps: Data standardization: The Z-score standardization method is used to standardize the values ​​of each variable; Calculate the covariance matrix: After data standardization, calculate the covariance matrix for historical survey data; Solving for eigenvalues ​​and eigenvectors: Expanding the covariance matrix to find the eigenvalues ​​and eigenvectors; Principal component selection: Principal components are selected based on the solved eigenvalues ​​and eigenvectors; Calculate the data score: Calculate the data score using the selected principal components and standardized historical survey data.

5. The photovoltaic project survey data integration and analysis system of claim 1, wherein: The method of using an artificial intelligence model to learn the correlation between semantic analysis, image features, and data scores of historical survey data from different regions includes the following steps: Data integration and preparation: The semantic analysis results, image features, and data scores of historical survey data from different regions are comprehensively integrated and categorized by region; Model selection: Based on the characteristics and complexities of the survey data from different regions, select an artificial intelligence model; Model training: Using semantic parsing, image features, and data scores from different regions as input, the model learns the correlation patterns between data from different regions.

6. The photovoltaic project survey data integration and analysis system of claim 1, wherein: The deep learning model, configured as a multilayer perceptron, includes the following construction steps: Determine the input layer: Determine the number of neurons in the input layer based on the dimensions of the survey data, and use the raw data as input to provide basic information for model learning; Designing hidden layers: The number of hidden layers and the number of neurons in each layer are determined through experiments and experience. Layers are connected by weight matrices. Each neuron receives the output of the neuron in the previous layer through a weighted summation and undergoes a nonlinear transformation through an activation function. Constructing the output layer: The output layer has one neuron that receives the output of the last hidden layer, and the predicted value is obtained by weighted summation and a linear activation function; Initialize weights and biases: After the model is built, the weight matrices of each layer and the biases of the neurons are initialized. The initialization is done by randomly initializing the weights. Select loss function and optimizer: select mean square error loss function to measure the difference between prediction and actual power generation, and use stochastic gradient descent method to adjust the weight bias to minimize the loss value.

Images