Photovoltaic module cleanliness evaluation method and system based on multi-dimensional data fusion

Abstract

The invention discloses a photovoltaic module cleanliness evaluation method and system based on multi-dimensional data fusion. The method comprises the following steps: acquiring multi-dimensional data of a photovoltaic power station, including generating capacity data, power station state data, irradiation data, meteorological data, geographic position data, cleaning record data, weather forecast data and economic index data; preprocessing the data; constructing a multi-dimensional feature; a pre-trained Transform architecture time sequence model is adopted to carry out training; evaluating the cleanliness state of the photovoltaic module in real time and predicting the change trend; and determining the optimal cleaning time and generating decision suggestions based on the cleanliness evaluation result and economic analysis. Through multi-dimensional data fusion and a large model algorithm, precise evaluation and scientific decision making of the cleanliness of the photovoltaic module are realized, the evaluation accuracy is improved by 15-20%, and the method is suitable for photovoltaic power stations at different geographic positions and under different climate conditions.

Description

Technical Field

[0001] This application relates to the technical field of photovoltaic power generation, and specifically relates to a method and system for evaluating the cleanliness of photovoltaic modules based on multi-dimensional data fusion. Background Art

[0002] As an important part of clean energy, the power generation efficiency of photovoltaic power generation directly affects the economic benefits of the power generation system. Dust accumulation on the surface of photovoltaic modules will significantly reduce the power generation efficiency. Therefore, timely and reasonable cleaning and maintenance are crucial. Currently, most photovoltaic power stations still adopt traditional regular cleaning methods, which have the following problems:

[0003] (1) The setting of the cleaning cycle lacks a scientific basis and is often based on empirical estimates; (2) The influence of factors such as geographical location and seasonal changes on the cleaning requirements is not considered; (3) It is impossible to make dynamic adjustments according to the actual dirt level; (4) It is difficult to achieve an optimal balance between the cleaning cost and the benefit.

[0004] To solve the above problems, various methods for evaluating the cleanliness of photovoltaic modules have been developed in the industry, but these methods still have limitations to varying degrees. Patent (CN117787551A) discloses a method for making a cleaning decision for photovoltaic modules based on photovoltaic power generation and light irradiance. This method uses the existing monitoring system of the power station, without the need to install additional sensors, and evaluates the cleanliness of the modules by analyzing the cumulative data of power generation and irradiance. However, this method does not take important factors such as seasonal changes and geographical location into consideration, resulting in obvious limitations in the accuracy of the evaluation results.

[0005] Patent (CN118095816A) proposes a cleanliness evaluation scheme based on a contrast plate from a different perspective. This method sets two photovoltaic panels, one equipped with an automatic cleaning system and the other not, and then judges the cleanliness by comparing the change in the power generation ratio of the two. The core assumption of this method is that the cleanliness of the on-site modules is consistent with that of the contrast plate, but the actual operating environment is much more complex than the experimental design. Differences in the installation angles of on-site modules, changes in the operating conditions of inverters, and uneven dust accumulation of modules in different regions will all cause deviations between the evaluation results and the actual situation.

[0006] Another approach can be seen in Patent (CN117436837A). This method judges the cleanliness based on the ratio of irradiance to power generation, and combines the installed capacity of the site, photovoltaic electricity price, and cleaning cost for economic analysis. Although this method considers the economy of the cleaning decision, the irradiance and power used are instantaneous values. Compared with using cumulative values such as power generation and cumulative irradiance, instantaneous values are more affected by the environment and have stronger volatility, resulting in insufficient stability of the evaluation results. In addition, this method does not fully consider the influence of external factors such as weather and seasons.

[0007] Patent (CN116596517A) attempts to apply artificial intelligence technology and proposes a cleaning decision-making method for photovoltaic modules based on a neural network model. This method mainly relies on long short-term memory networks for cleaning cycle prediction, but there are obvious deficiencies in practical applications: it fails to effectively combine the prediction results with actual operation and maintenance decisions, does not establish a comprehensive trade-off mechanism for cleaning benefits and costs, and does not consider the impact of expected factors such as weather forecasts on decisions. Therefore, it does not yet have the integrity to serve as a decision-making basis.

[0008] In summary, there are still many common problems in the existing technology for evaluating the cleanliness of photovoltaic modules: the evaluation indicators are single, and multi-dimensional data resources are not fully utilized; the impacts of geographical location and seasonal changes are not adequately considered; there is a lack of in-depth adaptation to actual operation and maintenance scenarios; and the scientific nature and systematicness of the decision-making process need to be improved. These problems directly affect the operation and maintenance efficiency and economic benefits of photovoltaic power plants, and there is an urgent need for a more comprehensive and accurate cleanliness evaluation method. Summary of the Invention

[0009] The present invention aims to solve the problems existing in the prior art and provides a method and system for evaluating the cleanliness of photovoltaic modules based on multi-dimensional data fusion. The present invention realizes the accurate evaluation of the cleanliness of photovoltaic modules and scientific cleaning decision-making by collecting multi-dimensional data, constructing a data fusion model, and combining large model algorithms, thereby improving power generation efficiency and optimizing operation and maintenance costs.

[0010] In a first aspect, an embodiment of the present application provides a method for evaluating the cleanliness of photovoltaic modules based on multi-dimensional data fusion, the method comprising:

[0011] A data acquisition step of collecting multi-dimensional data of a photovoltaic power plant, the multi-dimensional data including: power generation data, power plant status data, irradiation data, meteorological data, geographical location data, cleaning record data, weather forecast data, and economic indicator data;

[0012] A data preprocessing step of cleaning, screening, standardizing, and time-aligning the collected raw data to extract high-quality training data;

[0013] A feature engineering step of constructing multi-dimensional features based on the preprocessed data, including time features, geographical features, meteorological features, and power generation features, to form a comprehensive feature vector;

[0014] A model training step of selecting an appropriate machine learning model and combining it with a large model architecture to train a cleanliness evaluation model and establish a mapping relationship between the power plant operation status and cleanliness;

[0015] A cleanliness evaluation step of using the trained model to real-time evaluate the cleanliness status of photovoltaic modules and predict the future cleanliness change trend;

[0016] The cleaning decision-making step determines the optimal cleaning time based on the cleanliness evaluation results and economic analysis, and generates cleaning decision-making suggestions.

[0017] Optionally, in an implementation manner of the first aspect of the present invention, in the data collection step, the specifically collected data includes:

[0018] Power generation data: the daily power generation of the photovoltaic power station;

[0019] Power station status data: including maintenance status, power curtailment status, fault status, and cleaning status;

[0020] Irradiation data: the daily cumulative irradiation data of the irradiator in the photovoltaic power station;

[0021] Meteorological data: including daily average temperature, daily maximum temperature, daily minimum temperature, daily average humidity, daily precipitation, wind speed, and wind direction; Geographic location data: including longitude, latitude, altitude, and terrain type;

[0022] Cleaning record data: including manual cleaning date and heavy rain date;

[0023] Weather forecast data: including forecast precipitation, forecast temperature, forecast humidity, and forecast wind speed;

[0024] Economic data: including electricity price, unit cleaning cost, installed capacity, component area, and labor cost.

[0025] Optionally, in an implementation manner of the first aspect of the present invention, the data preprocessing step includes:

[0026] Data cleaning and screening: screening high-quality training samples, detecting and processing outliers, filling in missing values, and filtering and smoothing data;

[0027] Data standardization: performing normalization processing on data in different dimensions;

[0028] Time alignment: ensuring the consistency of timestamps of various types of data;

[0029] Data quality assessment: evaluating data integrity and consistency, and grading and marking data reliability.

[0030] Optionally, in an implementation manner of the first aspect of the present invention, the data cleaning and screening step specifically includes:

[0031] High-quality data identification: preferentially selecting data segments after cleaning the photovoltaic modules or after heavy rain as standard training data;

[0032] Benchmark establishment: calculating the ratio of power generation to cumulative irradiation in the clean state as the evaluation benchmark;

[0033] Operating status filtering: Exclude data when the power station is in a faulty, under maintenance, or power rationing state;

[0034] Physical constraint verification: Exclude data that does not conform to physical laws;

[0035] Sensor fault identification: Identify and mark sensor fault data;

[0036] Statistical anomaly detection: Use statistical methods to identify outliers;

[0037] Seasonal correction: Establish seasonal correction coefficients according to the power generation characteristics of each season.

[0038] Optionally, in an implementation manner of the first aspect of the present invention, the features constructed in the feature engineering step include:

[0039] Time features: Include seasonal features, month features, date features, and sunshine duration features;

[0040] Geographical features: Include location features, altitude features, component inclination features, and terrain features;

[0041] Meteorological features: Include temperature features, humidity features, precipitation features, wind force features, and weather comprehensive index;

[0042] Power generation features: Include power generation efficiency, irradiance response, power generation stability, and theoretical power generation.

[0043] Optionally, in an implementation manner of the first aspect of the present invention, the calculation formula for the theoretical power generation is:

[0044] E theory = η·A·H day ·f(T)·f(G),

[0045] where η is the component efficiency, A is the component area, H day is the daily irradiance cumulative amount, f(T) is the temperature correction function, and f(G) is the geographical location correction function.

[0046] Optionally, in an implementation manner of the first aspect of the present invention, the model training step includes:

[0047] Model selection: Use a pre-trained Transformer architecture time series model as the basic model;

[0048] Fine-tuning strategy: Adopt a hierarchical fine-tuning method, including two stages of freezing some levels and fine-tuning the entire model;

[0049] Model parameter setting: Set the input dimension, hidden layer dimension, number of attention heads, number of layers, learning rate, batch size, and number of training epochs of the model;

[0050] Model training process: Divide the dataset, design the loss function, select the optimizer and regularization method;

[0051] Domain adaptation: Use data in the photovoltaic domain for continuous pre-training;

[0052] Knowledge distillation: Use the integrated results of multiple models as soft labels to guide the learning of a single model;

[0053] Model evaluation and selection: Select the optimal model based on the evaluation metrics.

[0054] Optionally, in an implementation manner of the first aspect of the present invention, the fine-tuning strategy specifically includes:

[0055] The first stage: Freeze the front part of the pre-trained model, only train the rear part and the task adaptation layer, and use a larger learning rate;

[0056] The second stage: Unfreeze all layers and perform full-model fine-tuning using a smaller learning rate;

[0057] Learning rate scheduling: Adopt the cosine annealing scheduling strategy;

[0058] Early stopping strategy: Stop training when the performance of the validation set no longer improves for multiple consecutive rounds.

[0059] Optionally, in an implementation manner of the first aspect of the present invention, the loss function includes:

[0060] Prediction error term: L pred = MSE(y pred , y true );

[0061] Temporal continuity constraint: L temp = ∑|y t - y t-1 |·λ t ;

[0062] Physical model constraint: L phys = |y pred - y phys |·λ p ;

[0063] Knowledge distillation loss: L distill = KL(softmax(z s / T), softmax(z t / T))·λ d ;

[0064] Total loss: L = L pred + L temp + L phys + L distill。

[0065] Optionally, in an implementation of the first aspect of the present invention, the cleanliness evaluation step includes:

[0066] Calculation of cleanliness index: CI = 1 - (E theory - E actual ) / E theory ;

[0067] Cleanliness grading: The cleanliness is divided into four grades: excellent (A), good (B), medium (C), and poor (D);

[0068] Prediction of cleanliness change trend: Predict the cleanliness change in the next 7 days;

[0069] Generation of cleanliness evaluation report: including the current cleanliness status, cleanliness change trend chart, analysis of influencing factors, and cleaning suggestion warning.

[0070] Optionally, in an implementation of the first aspect of the present invention, the cleaning decision step includes:

[0071] Cost-benefit analysis: Calculate the cleaning cost, estimate the power generation loss, calculate the cleaning benefit, and return on investment;

[0072] Decision optimization: Determine whether to clean and the optimal cleaning time to maximize the return on cleaning investment;

[0073] Decision output: Generate cleaning suggestions, optimal cleaning time, expected benefits, and explanations of decision-making basis.

[0074] In a second aspect, an embodiment of the present application provides a photovoltaic module cleanliness evaluation system based on multi-dimensional data fusion, including:

[0075] Data acquisition module: used to acquire multi-dimensional data of a photovoltaic power station;

[0076] Data preprocessing module: used to clean, screen, standardize, and time-align the acquired raw data;

[0077] Feature engineering module: used to construct multi-dimensional features to form a comprehensive feature vector;

[0078] Model training module: used to train a cleanliness evaluation model;

[0079] Cleanliness evaluation module: used to evaluate the cleanliness status of photovoltaic modules in real time;

[0080] Cleaning decision module: used to determine the optimal cleaning time and generate cleaning decision suggestions.

[0081] Optionally, in an implementation of the second aspect of the present invention, the system adopts a distributed architecture, including:

[0082] Data acquisition layer: responsible for obtaining data from multiple sources;

[0083] Data processing layer: cleaning, standardizing, and extracting features from the raw data;

[0084] Model calculation layer: running the cleanliness evaluation algorithm to generate cleanliness evaluation results;

[0085] Decision output layer: based on the evaluation results, executing the decision optimization algorithm to generate cleaning suggestions.

[0086] Optionally, in an implementation manner of the second aspect of the present invention, the data preprocessing module includes:

[0087] Data cleaning unit: used for screening high-quality training samples, detecting and processing outliers, filling missing values, and filtering and smoothing data;

[0088] Data standardization unit: used for normalizing data of different dimensions;

[0089] Time alignment unit: used to ensure the consistency of timestamps of various types of data;

[0090] Data quality evaluation unit: used for evaluating data integrity and consistency, and grading and marking data reliability.

[0091] Optionally, in an implementation manner of the second aspect of the present invention, the feature engineering module includes:

[0092] Time feature construction unit: used for constructing seasonal features, month features, date features, and sunshine duration features;

[0093] Geographical feature construction unit: used for constructing location features, altitude features, component inclination features, and terrain features;

[0094] Meteorological feature construction unit: used for constructing temperature features, humidity features, precipitation features, wind force features, and weather comprehensive index;

[0095] Power generation feature construction unit: used for constructing power generation efficiency, irradiation response, power generation stability, and theoretical power generation;

[0096] Feature fusion unit: used for weighted fusion of different types of features to form a comprehensive feature vector.

[0097] Optionally, in an implementation manner of the second aspect of the present invention, the cleaning decision module includes:

[0098] Cost calculation unit: used for calculating the cleaning cost;

[0099] Benefit evaluation unit: used to estimate power generation losses, calculate cleaning benefits and return on investment;

[0100] Decision optimization unit: used to determine whether to clean and the optimal cleaning time to maximize the return on cleaning investment;

[0101] Decision output unit: used to generate cleaning suggestions, optimal cleaning time, expected benefits and explanations of decision-making basis.

[0102] Optionally, in an implementation manner of the second aspect of the present invention, a visualization display module is further included, which is used for:

[0103] Display the real-time status of cleanliness;

[0104] Show the change trend of cleanliness;

[0105] Output a cleaning suggestion report;

[0106] Visualize the cost-benefit analysis results.

[0107] Optionally, in an implementation manner of the second aspect of the present invention, the model training module uses a pre-trained Transformer architecture time series model as the basic model and is trained through hierarchical fine-tuning and domain adaptation methods. The loss function of the model includes a prediction error term, a time continuity constraint, a physical model constraint, and a knowledge distillation loss.

[0108] The present invention realizes the accurate evaluation and scientific decision-making of the cleanliness of photovoltaic modules through multi-dimensional data fusion and large model algorithms, and has the following remarkable effects:

[0109] (1) Improved evaluation accuracy: Compared with existing similar evaluation methods (such as CN117787551A), the accuracy of cleanliness evaluation is increased by about 15-20%.

[0110] (1) More scientific cleaning decision-making: Based on multi-dimensional data and accurate models, the accurate timing of cleaning is realized, avoiding blind cleaning or cleaning delays.

[0111] (3) Maximized economic benefits: Although the cleaning frequency and related costs may increase in some cases, through accurate timing selection, the additional benefits brought by the increased power generation significantly exceed the increased cleaning costs, realizing the maximization of net benefits.

[0112] (4) Strong system adaptability: Applicable to photovoltaic power stations in different geographical locations and different climate conditions, and can dynamically adjust the decision-making model according to actual situations. Description of the Drawings

[0113] Figure 1 It is the system architecture diagram of an embodiment of the present invention.

[0114] Figure 2 The flowchart of feature engineering according to an embodiment of the present invention.

[0115] Figure 3 The flowchart of data processing according to an embodiment of the present invention.

[0116] Figure 4 The structural diagram of the cleanliness evaluation model according to an embodiment of the present invention.

[0117] Figure 5 The flowchart of decision optimization according to an embodiment of the present invention. Specific implementation manners

[0118] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments.

[0119] It should be noted that in the embodiments of the present application, "at least one" means one or more, and "a plurality" means two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present application belongs. The terms used in the specification of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application.

[0120] It should be noted that in the embodiments of the present application, terms such as "first" and "second" are only used for the purpose of distinguishing descriptions, and cannot be understood as indicating or implying relative importance, nor can they be understood as indicating or implying order. Features defined with "first" and "second" can explicitly or implicitly include one or more of the described features. In the description of the embodiments of the present application, words such as "exemplary" or "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the embodiments of the present application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of words such as "exemplary" or "for example" is intended to present relevant concepts in a specific manner.

[0121] Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts are within the scope of protection of the present application.

[0122] Embodiment 1

[0123] Overall system architecture and working process

[0124] The system of the present invention adopts a distributed architecture, including four main components: a data acquisition layer, a data processing layer, a model calculation layer, and a decision output layer. The system architecture is as Figure 1 shown.

[0125] The functions and interaction methods of each layer are as follows:

[0126] 1. Data acquisition layer: Responsible for obtaining data from multiple sources, collecting data through standardized API interfaces or database connection methods, mainly including power station monitoring systems, weather stations, power trading platforms, and geographic information systems, etc.

[0127] 2. Data processing layer: Cleans, standardizes, and extracts features from the original data, and uses a distributed computing framework to improve processing efficiency.

[0128] 3. Model calculation layer: Runs a cleanliness evaluation algorithm, performs inference based on a large model architecture, and generates a cleanliness evaluation result.

[0129] 4. Decision output layer: Executes a decision optimization algorithm based on the evaluation result, generates cleaning suggestions, and displays them to the user through a visual interface.

[0130] The system working process is: data acquisition → data preprocessing → feature engineering → model training / inference → cleanliness evaluation → cleaning decision → result display. The entire process forms a closed loop, and the system will continuously collect new data to optimize the model performance.

[0131] Embodiment 2

[0132] Data acquisition and preprocessing solution

[0133] 2.1 Data acquisition module

[0134] This module is responsible for collecting data related to cleanliness evaluation from multiple sources, specifically including:

[0135] 2.1.1 Power generation data: Collect the daily power generation E day (kWh) of the photovoltaic power station, which is used as a basic indicator for evaluating cleanliness.

[0136] 2.1.2 Power station status data: Collect the operating status S of the power station, including:

[0137] Maintenance status S m : The value range is {0, 1}, 1 indicates in maintenance;

[0138] Power rationing status S l : The value range is {0, 1}, 1 indicates in power rationing;

[0139] Fault status S f : The value range is {0, 1}, 1 indicates there is a fault;

[0140] Power rationing status S c : The value range is {0, 1}, 1 indicates already cleaned;

[0141] 2.1.3 Irradiation data: Collect the daily cumulative irradiation data of the irradiators in the PV power station

[0142] H day (kWh / m 2 ), as the reference benchmark for power generation.

[0143] 2.1.4 Meteorological data: Obtain the meteorological information of the site location from the meteorological department, including:

[0144] Average sunshine temperature T avg (°C), daily maximum temperature T max (°C), daily minimum temperature T min (°C), daily average humidity Hum avg (%), daily precipitation Rain (mm), wind speed Wind v (m / S) and wind direction Wind d (°).

[0145] 2.1.5 Geographic location data: Record the geographical information of the power station, including: longitude Lon (°), latitude Lat (°), altitude Elev (m), terrain type Terrain type .

[0146] 2.1.6 Cleaning record data: Record the cleaning information of the power station, including: manual cleaning date C date , which is also the latest cleaning date Last used for decision-making in the system clean ; date of heavy rain (>50mm) R date , as a mark of natural cleaning.

[0147] 2.1.7 Weather forecast data: Obtain the weather forecast information for the next 7 - 15 days, including: forecast precipitation Rain forecast (mm), forecast temperature T forecast (°C), forecast humidity Hum forecast (%), forecast wind speed Wind v,forecast (m / S)(m / S), weather suitability determination

[0148] Weather suitable , taking values {0, 1}, where 1 indicates suitable for cleaning.

[0149] 2.1.8 Economic data: Collect economic indicators related to cleaning decision-making, including: electricity price Price (yuan / kWh), unit cleaning cost Cost clean (yuan / m 2 ), installed capacity Capacity (kW), module area Area (m 2 ), labor cost Labor cost (yuan / time).

[0150] 2.2 Data preprocessing module

[0151] This module cleans, standardizes, and aligns the collected raw data to ensure the accuracy of subsequent analysis.

[0152] 2.2.1 Data cleaning and screening:

[0153] Basic data screening: Screen valid training samples from the data collected from dozens of photovoltaic power plants across the country.

[0154] High-quality data identification: Prioritize selecting data segments (3 - 10 days) after cleaning the photovoltaic modules or after heavy rain (>50 mm) as standard training data, when the system is in an ideal power generation state.

[0155] Benchmark establishment: Calculate the ratio R of the power generation under clean conditions to the cumulative irradiation clean = E day / H day , as the evaluation benchmark.

[0156] Operating status filtering: Eliminate data when the power plant is in a fault state (s f = 1), maintenance state (s m = 1), or power curtailment state (s l = 1).

[0157] Outlier detection and handling:

[0158] Physical constraint check: Eliminate data that does not conform to physical laws, such as negative power generation (E day < 0) or extremely high values E day > 1.2·Capacity·2.4).

[0159] Sensor fault identification: Identify and mark data with faulty irradiance sensors (H day = 0 and clear weather).

[0160] Statistical outlier detection: Use the 3σ criterion to identify outliers, that is, for data x, if |x - μ| > 3σ, then it is determined as an outlier (μ is the mean, σ is the standard deviation).

[0161] Seasonal correction: Establish a seasonal correction coefficient α season according to the power generation characteristics of each season, and correct the missing value of the outlier detection threshold. For time series data, use the interpolation algorithm to fill in the missing values.

[0162] Short-term missing data (≤ 3 days): Use linear interpolation.

[0163] Medium-term missing data (4 - 7 days): Use polynomial interpolation.

[0164] Long-term absence (>7 days): The similar-day pattern replacement method is adopted, specifically:

[0165] a. Select similar dates in history based on date features (day of the week, season);

[0166] b. Calculate the similarity score Sim score = w1·Sim season + w2·Sim weather + w3·Sim dayofweek ;

[0167] c. Select the date data with the highest similarity for replacement.

[0168] 2.2.3 Time alignment: Ensure the consistency of timestamps for various types of data.

[0169] Uniformly adopt the UTC+8 time zone

[0170] Align the data to natural days (00:00 - 24:00)

[0171] For data with different sampling frequencies, uniformly aggregate them to daily granularity.

[0172] 2.2.4 Data quality assessment:

[0173] Completeness score: Calculate the completeness rate Completeness of data in each dimension = N valid / N total .

[0174] Consistency check: Check the consistency between different data sources of the same power station.

[0175] Reliability marking: Classify and mark the data reliability as the weight basis for model training.

[0176] After data preprocessing, a high-quality dataset D clean = {d i ' ,j} is obtained, where each record contains complete time stamps, power station status, power generation, irradiation amount and other information, providing a reliable basis for subsequent model training.

[0177] Example 3:

[0178] Feature engineering and model training scheme

[0179] 3.1 Feature engineering module

[0180] Based on the preprocessed data, this module constructs multi-dimensional features to form a comprehensive feature vector. The feature engineering process is as Figure 2 shown.

[0181] 3.1.1 Temporal Feature Construction:

[0182] Date Encoding T date : Convert the date into continuous numerical values

[0183] Season Feature T season : Spring (1), Summer (2), Autumn (3), Winter (4)

[0184] Month Feature T month : 1 - 12

[0185] Week Feature T week : 1 - 52

[0186] Day Feature T day : 1 - 31

[0187] Day of the Week Feature T weekday : 1 - 7

[0188] Holiday Feature T holiday : Whether it is a public holiday, taking values {0, 1}

[0189] Sunshine Duration Index D i : The theoretical sunshine duration calculated based on astronomical calculations, and the calculation formula is:

[0190] where φ is the latitude and δ is the solar declination angle.

[0191] 3.1.2 Geographic Feature Construction:

[0192] Latitude and Longitude Encoding G loc : Convert the latitude and longitude into spherical coordinates

[0193] Elevation Stratification G elev : Divide the elevation into three layers: low (1), medium (2), and high (3)

[0194] Terrain Feature G terrain : Plain (1), Hilly (2), Mountainous (3), Plateau (4), etc.

[0195] Dust Index G dust : The regional dust accumulation rate calculated based on historical data

[0196] 3.1.3 Meteorological Feature Construction:

[0197] Temperature Feature W temp : Include daily average, maximum, minimum temperature and temperature difference

[0198] Humidity Feature W hum : Daily average humidity and change trend

[0199] Precipitation Feature W rain: Daily precipitation and precipitation frequency

[0200] Wind speed characteristic W wind : Daily average wind speed and maximum wind speed

[0201] Wind direction coefficient W wind_dir : Coefficient calculated according to the wind direction, the formula is:

[0202] W wind_dir = sin(θ - α)+1, where θ is the wind direction angle and α is the power station azimuth angle.

[0203] Comprehensive weather index W index : Weighted sum of multiple meteorological factors, the calculation formula is:

[0204] W index = w1W temp + w2W hum + w3W rain + w4W wind , where the weight coefficients are adjusted according to the characteristics of different regions.

[0205] 3.1.4 Power generation characteristic construction:

[0206] Power generation efficiency P eff : Ratio of daily power generation to the theoretical maximum power generation

[0207] Irradiance response P resp : Slope of the response curve of power generation to irradiance

[0208] Power generation stability P stab : Coefficient of variation of power generation

[0209] Theoretical power generation E theory : Calculated from the theoretical model established based on historical data under clean conditions, the calculation formula is:

[0210] E theory = η·A·H day ·f(T)·f(G),

[0211] where η is the module efficiency, A is the module area, f(T) is the temperature correction function, and f(G) is the geographical location correction function.

[0212] 3.1.5 Feature selection and fusion:

[0213] Feature importance assessment: Calculate the feature importance score using the random forest algorithm.

[0214] Multi-feature fusion: Weightedly fuse different types of features to obtain a comprehensive feature vector.

[0215] Feature Dimension Optimization: Reduce the feature dimension through Principal Component Analysis (PCA) to improve computational efficiency.

[0216] After feature engineering, the feature vector matrix X = {x i,j} and the label vector Y = {y i} are obtained as the input for model training.

[0217] 3.2 Model Training Module

[0218] This module trains the cleanliness evaluation model based on the constructed features. The model structure is as Figure 4 shown.

[0219] 3.2.1 Model Selection:

[0220] Adopt the pre-trained large model of the Transformer architecture as the base model, and specifically select the Time Series Transformer (TST) of Hugging Face as the pre-trained model.

[0221] This model has been pre-trained on a large amount of time series data and has a strong ability to extract time series features.

[0222] Comparative experiments show that the fine-tuning method based on the pre-trained model has the following advantages compared with models such as LSTM, GRU, random forest, and XGBoost trained from scratch:

[0223] Faster convergence speed: The number of training rounds required is reduced by 60%.

[0224] Lower sample requirements: Only 40% of the original data volume is required to achieve the same performance.

[0225] Higher prediction accuracy: The prediction error is reduced by about 25%.

[0226] 3.2.2 Model Architecture Customization:

[0227] Base architecture: The pre-trained TST model (6-layer Transformer encoder).

[0228] Input adaptation layer: Map the multi-dimensional feature vector into the model input format.

[0229] Time series encoding: Position encoding + time feature encoding.

[0230] Feature embedding: Convert different types of features into embedding vectors of a unified dimension.

[0231] Attention mechanism: Multi-head self-attention mechanism (8 heads), effectively capturing the dependencies between different features.

[0232] Task Adaptation Layer: Add an output layer specific to cleanliness assessment.

[0233] Fully Connected Layers: 128 dimensions → 64 dimensions → 32 dimensions → 1 dimension.

[0234] Activation Functions: A combination of ReLU and Sigmoid.

[0235] 3.2.3 Fine-tuning Strategy:

[0236] Hierarchical Fine-tuning: Adopt a fine-tuning strategy of unfreezing layer by layer

[0237] First Stage: Freeze the first 4 layers of the Transformer encoder and only train the last 2 layers and the task adaptation layer

[0238] Second Stage: Unfreeze all layers and perform full-model fine-tuning with a smaller learning rate

[0239] Learning Rate Setting:

[0240] Initial Learning Rate: 0.0001

[0241] Decay Strategy: Cosine annealing schedule with a minimum learning rate of 0.000001

[0242] Batch Size: 64

[0243] Number of Training Epochs: 50 epochs in the first stage and 150 epochs in the second stage. Adopt an early stopping strategy (no improvement in the validation set for 10 epochs).

[0244] 3.2.4 Model Training Process:

[0245] Dataset Partition: Training Set (70%), Validation Set (15%), Test Set (15%)

[0246] Loss Function Design:

[0247] Prediction Error Term: L pred = MSE(y pred , y true )

[0248] Temporal Continuity Constraint: L temp = ∑|y t - y t-1 | · λ t , where λ t = 0.2

[0249] Physical Model Constraint: L phys = |y pred - y phys | · λ p , where λ p = 0.5

[0250] Total loss: L = L pred + L temp + L phys

[0251] Optimizer: AdamW optimizer, weight decay coefficient 0.01

[0252] Regularization: Use Dropout(0.2) and gradient clipping (maximum range 1.0)

[0253] 3.2.5 Domain Adaptation and Knowledge Distillation:

[0254] Domain Adaptation: Adapt the pre-trained model to the photovoltaic power generation field, introduce the domain vocabulary: photovoltaic technical terms and concepts, domain pre-training: use unlabeled photovoltaic time series data for continued pre-training.

[0255] Knowledge Distillation: Use the results of multiple model integrations as "soft labels". Teacher model: An integration of 5 TST models initialized with different parameters, distillation temperature: 2.0, distillation loss weight: 0.3.

[0256] 3.2.6 Model Evaluation and Selection:

[0257] Evaluation metrics: RMSE, MAE, R 2

[0258] Model selection criterion: Optimal performance on the validation set

[0259] Post-processing strategy: Calibrate the model output using feedback data

[0260] 3.2.7 Handling of Model Boundary Conditions:

[0261] Handling of extreme meteorological conditions: When the meteorological conditions exceed the range of the training data, a correction method based on a physical model is adopted. Handling of newly installed power stations: For new power stations without historical data, a transfer learning method using data from similar power stations is adopted. Adaptation to seasonal changes: Update the model parameters quarterly to ensure that the model adapts to seasonal changes.

[0262] After fine-tuning, the final cleanliness evaluation model M is obtained fine-tuned , the total number of parameters of the model is approximately 32.5MB, and the inference time is less than 100ms / sample.

[0263] Example 4

[0264] Cleanliness Evaluation and Decision Making Implementation

[0265] 4.1 Cleanliness Evaluation Module

[0266] This module uses the trained model to evaluate the cleanliness status of components in real time.

[0267] 4.1.1 Calculation of cleanliness index:

[0268] Calculation formula for cleanliness index CI:

[0269] CI = 1 - (E theory - E actual ) / E theory

[0270] Cleanliness grading:

[0271] Excellent (A): CI ≥ 0.95

[0272] Good (B): 0.90 ≤ CI ≤ 0.95

[0273] Medium (C): 0.85 ≤ CI ≤ 0.90

[0274] Poor (D): CI ≤ 0.85

[0275] 4.1.2 Prediction of cleanliness change trend:

[0276] Time window: The next 7 days

[0277] Input of prediction model: [X current , Weather forecast

[0278] Prediction output: CI pred = [CI t+1 , CI t+2 ,..., CI t+7

[0279] Calculation of change rate: Rate = (CI t+7 - CI t ) / 7

[0280] 4.1.3 Generation of cleanliness assessment report:

[0281] Current cleanliness status

[0282] Cleanliness change trend chart

[0283] Analysis of influencing factors

[0284] Warning of cleaning suggestions

[0285] 4.2 Cleaning decision-making module

[0286] Based on the cleanliness assessment results, this module conducts cleaning decision-making analysis. The decision optimization process is as Figure 5 shown.

[0287] 4.2.1 Cost-benefit analysis:

[0288] Calculation of cleaning cost: C clean ​​= Area·Cost clean + Labor cost

[0289] Estimation of power generation loss: Loss = ∑(1 - CI i )·E theoryi ·Price

[0290] Calculation of cleaning benefit: Benefit = Loss - C clean

[0291] Return on investment: ROI = Benefit / C clean

[0292] 4.2.2 Decision optimization:

[0293] Decision variables: Whether to clean (0 / 1), cleaning time t

[0294] Objective function: Maximize the cleaning ROI

[0295] Constraints:

[0296] Weather suitability: Weather suitable (t) = 1, determined based on the following conditions:

[0297] a. Forecast precipitation probability < 30%

[0298] b. Forecast wind speed < 5 m / s

[0299] c. Forecast temperature > 5 °C

[0300] Power station operating status: S op (t) = 1

[0301] Minimum cleaning interval: t - Last clean ≥ Min interval where Min interval = 30 days 4.23 Decision output:

[0302] Cleaning recommendation: Clean suggest = {0, 1}

[0303] Optimal cleaning time: T optimal

[0304] Expected benefit: Expected benefit

[0305] Explanation of decision-making basis

[0306] Example 5

[0307] Analysis of typical application cases

[0308] This section demonstrates the application effect of the present invention in the actual environment through typical application cases. 5.1 Basic information of the power station:

[0309] Location: An industrial park in Zhejiang Province

[0310] Installed capacity: 3.5 MW

[0311] Component type: Monocrystalline silicon

[0312] Installation method: Roof fixed

[0313] Component inclination: 25°

[0314] Existing operation and maintenance strategy: Clean 3 times a year, and the contract is included in the contracting fee of the operation and maintenance company. 5.2 Application process:

[0315] Data collection: Connect the irradiance meter and infrared meter reading data of the power station, and collect the historical operation data for one year

[0316] Model training: Train the cleanliness evaluation model based on historical data

[0317] Real-time monitoring: Deploy the model for a 1-year cleanliness evaluation

[0318] Decision execution: Make cleaning decisions according to the system suggestions

[0319] 5.3 Implementation effect:

[0320] Accuracy rate of cleanliness evaluation: 92.5%

[0321] Optimization of cleaning times: Increased from the original 3 times / year to 6 times / year, and optimized the cleaning timing

[0322] Increase in power generation: The annual power generation increased by 4.8%

[0323] Economic benefit analysis:

[0324] Extra cleaning cost: 3 times × 3.5 MW × 2000 yuan / MW = 21,000 yuan

[0325] Power increase: 3.5 MW × 1200 kWh / kW / year × 4.8% = 201,600 kWh

[0326] This project is for self-use of electricity generated and selling the surplus electricity to the grid, with a relatively high self-use ratio. Calculated at a comprehensive electricity price of 0.62 yuan / kWh, the increased income is: 124,992 yuan

[0327] Net income increase: 103,992 yuan / year

[0328] 5.4 Comparative analysis:

[0329] Compared with the traditional regular cleaning method: Although the number of cleanings and costs increase, by optimizing the cleaning timing, the marginal benefit of each cleaning is higher.

[0330] Compared with the original experience-based decision-making: The cleanliness assessment is more accurate, the decision-making is more scientific, and the return on investment is significantly improved.

[0331] Return on investment: The ROI of the additional cleaning investment is approximately 495%, much higher than the industry average.

[0332] Industrial applicability

[0333] The present invention has been field-tested in multiple different types of photovoltaic power stations with obvious effects:

[0334] Accuracy rate of cleanliness assessment: On average, it reaches over 90%, an increase of approximately 15 - 20% compared with the existing assessment method;

[0335] Economic benefit optimization: Although the cleaning frequency and related costs usually increase, through scientific timing selection, the comprehensive economic benefit is increased by an average of 4 - 5%. For power stations with serious dust accumulation and a cleaning frequency far lower than the optimal level, the increase can reach 8 - 10%;

[0336] Improvement of power generation efficiency: Through timely and reasonable cleaning, the annual power generation of the power station is increased by an average of 3 - 5%.

[0337] The system integration of the present invention has significant economic advantages:

[0338] 1. Hardware transformation requirements:

[0339] Utilize standardized market general equipment to avoid the high costs brought by specialized hardware

[0340] Adopt modular design and can be selectively deployed according to the actual needs of the power station

[0341] Fully compatible with the existing monitoring system to minimize additional equipment investment

[0342] 2. Economic benefits:

[0343] Fast investment recovery speed: The investment payback period of a typical site is less than one quarter

[0344] High benefit multiple: The annual gain brought by the system is usually several times to dozens of times the initial investment

[0345] Obvious scale effect: The larger the site scale, the higher the return on unit investment

[0346] The present invention is applicable to various photovoltaic power stations, with relatively low hardware input costs, can be integrated with existing monitoring systems, and has high popularization and application value. It is especially suitable for industrial and commercial distributed and large-scale ground power stations, with a short investment recovery period and significant long-term economic benefits.

Claims

1. A method for evaluating the cleanliness of photovoltaic modules based on multi-dimensional data fusion, characterized in that, The method includes the following steps: Data collection step: Collect multi-dimensional data of the photovoltaic power station. The multi-dimensional data includes power generation data, power station status data, irradiation data, meteorological data, geographical location data, cleaning record data, weather forecast data, and economic index data; Data preprocessing step: Clean, screen, standardize, and time-align the collected raw data to extract high-quality training data; Feature engineering step: Based on the preprocessed data, construct multi-dimensional features, including time features, geographical features, meteorological features, and power generation features, to form a comprehensive feature vector; Model training step: Select an appropriate machine learning model and combine it with a large model architecture to train a cleanliness evaluation model and establish a mapping relationship between the power station operation status and cleanliness; Cleanliness evaluation step: Use the trained model to evaluate the cleanliness status of the photovoltaic modules in real time and predict the future cleanliness change trend; Cleaning decision-making step: Based on the cleanliness evaluation results and economic analysis, determine the optimal cleaning time and generate cleaning decision-making suggestions.

2. The method for evaluating the cleanliness of photovoltaic modules based on multi-dimensional data fusion according to claim 1, wherein In the data collection step, the specifically collected data includes: Power generation data: The daily power generation of the photovoltaic power station; Power station status data: including maintenance status, power curtailment status, fault status, and cleaning status; Irradiation data: The daily cumulative irradiation data of the irradiator in the photovoltaic power station; Meteorological data: including daily average temperature, daily maximum temperature, daily minimum temperature, daily average humidity, daily precipitation, wind speed, and wind direction; Geographical location data: including longitude, latitude, altitude, and terrain type; Cleaning record data: including the date of manual cleaning and the date of heavy rain; Weather forecast data: including forecast precipitation, forecast temperature, forecast humidity, and forecast wind speed; Economic data: including electricity price, unit cleaning cost, installed capacity, module area, and labor cost; The data preprocessing step includes: Data cleaning and screening: Screen high-quality training samples, detect and process outliers, fill in missing values, and filter and smooth the data; Data standardization: Normalize data of different dimensions; Time alignment: Ensure the consistency of timestamps of various types of data; Data quality evaluation: Evaluate data integrity and consistency, and grade and mark data reliability; The data cleaning and screening step specifically includes: Identification of high-quality data: Prioritize selecting data segments after cleaning the photovoltaic modules or after heavy rain as standard training data; Benchmark establishment: Calculate the ratio of power generation to cumulative irradiation under the clean state as the evaluation benchmark; Operation status filtering: Eliminate data when the power station is in a fault, maintenance, or power curtailment state; Physical constraint inspection: Eliminate data that does not conform to physical laws; Sensor fault identification: Identify and mark sensor fault data; Statistical anomaly detection: Use statistical methods to identify outliers; Seasonal correction: Establish a seasonal correction coefficient according to the power generation characteristics of each season.

3. The method for evaluating the cleanliness of photovoltaic modules based on multi-dimensional data fusion according to claim 1, wherein The features constructed in the feature engineering step include: Temporal features: including seasonal features, monthly features, daily features, and sunshine duration features; Geographical features: including location features, altitude features, component tilt angle features, and terrain features; Meteorological features: including temperature features, humidity features, precipitation features, wind force features, and weather composite index; Power generation features: including power generation efficiency, irradiance response, power generation stability, and theoretical power generation; The calculation formula for the theoretical power generation is: E theory = η·A·H day ·f(T)·f(G), where η is the component efficiency, A is the component area, H day is the cumulative daily solar irradiance, f(T) is the temperature correction function, and f(G) is the geographical location correction function.

5. The method for evaluating the cleanliness of a photovoltaic module based on multi-dimensional data fusion according to claim 1, wherein: The model training steps include: Model selection: Using a pre-trained Transformer architecture time series model as the basic model; Fine-tuning strategy: Adopting a hierarchical fine-tuning method, including two stages: freezing some levels and fine-tuning the entire model; Model parameter setting: Setting the input dimension, hidden layer dimension, number of attention heads, number of layers, learning rate, batch size, and number of training epochs of the model; Model training process: Dividing the dataset, designing the loss function, selecting the optimizer and regularization method; Domain adaptation: Using data in the photovoltaic domain for continuous pre-training; Knowledge distillation: Using the integrated results of multiple models as soft labels to guide the learning of a single model; Model evaluation and selection: Selecting the optimal model based on evaluation metrics; The fine-tuning strategy specifically includes: The first stage: Freezing the front part of the pre-trained model, only training the latter part and the task adaptation layer, and using a larger learning rate; The second stage: Thawing all layers and fine-tuning the entire model with a smaller learning rate; Learning rate scheduling: Adopting a cosine annealing scheduling strategy; Early stopping strategy: Stopping training when the performance of the validation set no longer improves for multiple consecutive rounds; The loss function includes: Prediction error term: L pred = MSE(y pred , y true ); Temporal continuity constraint: L temp = ∑|y t - y t-1 | · λ t ; Physical model constraint: L phys = |y pred - y phys |·λ p ; Knowledge distillation loss: L distill = KL(softmax(z s / T), softmax(z t / T)) · λ d ; Total loss: L = L pred +L temp +L phys +L distill .

5. The method for evaluating the cleanliness of a photovoltaic module based on multi-dimensional data fusion according to claim 1, wherein The cleanliness evaluation steps include: Cleanliness index calculation: CI = 1 - (E theory - E actual ) / E theory ; Cleanliness grading: Dividing the cleanliness into four grades: excellent (A), good (B), medium (C), and poor (D); Prediction of the cleanliness change trend: Predicting the cleanliness change in the next 7 days; Generation of a cleanliness evaluation report: Including the current cleanliness status, cleanliness change trend chart, factor analysis, and cleaning recommendation warning; 6. The method for evaluating the cleanliness of a photovoltaic module based on multi-dimensional data fusion according to claim 1, wherein The cleaning decision steps include: Cost-benefit analysis: Calculating the cleaning cost, estimating the power generation loss, calculating the cleaning benefit, and return on investment; Decision optimization: Determining whether to clean and the optimal cleaning time to maximize the return on cleaning investment; Decision output: Generating cleaning recommendations, optimal cleaning time, expected benefits, and an explanation of the decision basis; 7. A photovoltaic module cleanliness evaluation system based on multi-dimensional data fusion, characterized in that, Including: Data acquisition module: Used to acquire multi-dimensional data of a photovoltaic power station; Data preprocessing module: Used to clean, screen, standardize, and time-align the acquired raw data; Feature engineering module: Used to construct multi-dimensional features to form a comprehensive feature vector; Model training module: Used to train the cleanliness evaluation model; Cleanliness evaluation module: Used to evaluate the cleanliness status of photovoltaic modules in real time; Cleaning decision module: Used to determine the optimal cleaning time and generate cleaning decision recommendations; 8. The system for evaluating the cleanliness of a photovoltaic module based on multi-dimensional data fusion according to claim 7, wherein: The system adopts a distributed architecture, including: Data acquisition layer: Responsible for obtaining data from multiple sources; Data processing layer: Clean, standardize, and extract features from the original data; Model calculation layer: Run the cleanliness evaluation algorithm to generate the cleanliness evaluation results; Decision output layer: Based on the evaluation results, execute the decision optimization algorithm to generate cleaning suggestions; The data preprocessing module includes: Data cleaning unit: Used to screen high-quality training samples, detect and process outliers, fill in missing values, and filter and smooth data; Data standardization unit: Used to normalize data of different dimensions; Time alignment unit: Used to ensure the consistency of timestamps of various types of data; Data quality assessment unit: Used to evaluate data integrity and consistency, and grade and mark data reliability.

9. The photovoltaic module cleanliness evaluation system based on multi-dimensional data fusion according to claim 7, wherein The feature engineering module includes: Time feature construction unit: Used to construct seasonal features, month features, date features, and sunshine duration features; Geographical feature construction unit: Used to construct location features, altitude features, component inclination features, and terrain features; Meteorological feature construction unit: Used to construct temperature features, humidity features, precipitation features, wind force features, and weather comprehensive index; Power generation feature construction unit: Used to construct power generation efficiency, irradiance response, power generation stability, and theoretical power generation; Feature fusion unit: Used to perform weighted fusion of different types of features to form a comprehensive feature vector; The cleaning decision module includes: Cost calculation unit: Used to calculate the cleaning cost; Benefit evaluation unit: Used to estimate power generation losses, calculate cleaning benefits, and return on investment; Decision optimization unit: Used to determine whether to clean and the optimal cleaning time to maximize the return on cleaning investment; Decision output unit: Used to generate cleaning suggestions, optimal cleaning time, expected benefits, and explanations of decision-making basis.

10. The photovoltaic module cleanliness evaluation system based on multi-dimensional data fusion according to claim 7, wherein It further includes a visualization display module for: Displaying the real-time status of cleanliness; Showing the change trend of cleanliness; Outputting a cleaning suggestion report; Visualizing the cost-benefit analysis results; The model training module uses a pre-trained Transformer architecture time series model as the basic model and is trained by hierarchical fine-tuning and domain adaptation methods. The loss function of the model includes a prediction error term, a time continuity constraint, a physical model constraint, and a knowledge distillation loss.

Images