A photovoltaic power generation efficiency evaluation optimization method and system based on an internet of things

Abstract

This invention relates to the field of photovoltaic power generation technology and discloses a photovoltaic power generation efficiency evaluation and optimization method and system based on the Internet of Things. Specifically, it includes: constructing a photovoltaic module loss prediction model, which is a multi-branch CNN-feature fusion-XGBoost end-to-end hybrid model; during inference and evaluation, calling the multi-branch CNN-feature fusion-XGBoost end-to-end hybrid model, inputting fused features to obtain the loss rate prediction value, and combining it with SHAP value analysis to output the core loss source, thus completing the quantitative attribution of low efficiency; matching the model prediction results with a constructed loss threshold-core factor-operation and maintenance suggestion rule base, and pushing the operation and maintenance suggestions to the operation and maintenance platform; the operation and maintenance platform receives the decision instructions and executes the operation, thus completing the evaluation and optimization process. This invention solves the problems of traditional operation and maintenance relying on experience for decision-making, lagging efficiency evaluation, and unclear loss source tracing.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation technology, and in particular to a method and system for evaluating and optimizing photovoltaic power generation efficiency based on the Internet of Things. Background Technology

[0002] The photovoltaic power generation industry has made significant progress in terms of technology, industrial scale, and application depth, with installed capacity continuously increasing and costs gradually decreasing. However, in the actual application of photovoltaic power generation, its power generation efficiency is affected by a series of issues such as the external environment, component characteristics, system power loss, inverter efficiency, and grid connection control.

[0003] Current photovoltaic power generation efficiency assessments mostly employ a two-factor model based on illumination and temperature, which has significant technical limitations: First, it fails to incorporate key influencing factors such as spectral matching and module aging, resulting in assessment errors exceeding 10%. For example, crystalline silicon modules utilize 80% of visible light but only 30% of ultraviolet light, a difference that traditional models cannot quantify. Second, loss localization is vague, only calculating the total loss rate (typically 20%-30% in older power plants), failing to distinguish between specific loss sources such as spectral mismatch loss, temperature loss, and shading loss. Third, the model's generalization ability is weak; when models trained on single-power plant data are applied to similar scenarios, the error rate climbs to over 15%, lacking cross-scenario adaptability. Summary of the Invention

[0004] Photovoltaic module operation data encompasses multiple data sources such as spectrum, temperature, and voltage. The differences in their dimensions and modal heterogeneity create barriers to data fusion technology. Meanwhile, module losses are dynamically coupled and affected by multiple factors such as light intensity, dust accumulation, ambient temperature, and module aging. There are nonlinear correlations and interactions between these factors, making it difficult for traditional single models (such as a single CNN or XGBoost) to fully capture loss patterns. They generally suffer from insufficient accuracy and weak generalization ability, failing to meet the needs of refined operation and maintenance of power plants. Based on this, this invention provides a photovoltaic power generation efficiency evaluation and optimization scheme based on the Internet of Things.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] An IoT-based method for evaluating and optimizing photovoltaic power generation efficiency includes the following steps:

[0007] Step 1: The IoT data acquisition layer is built to collect real-time spectral data, temperature field distribution, voltage fluctuation curves, environmental parameters, and operation and maintenance records of photovoltaic modules, and transmit them to the processing layer through the edge layer and transmission layer.

[0008] Step 2: Construct a photovoltaic module loss prediction model. This model is a multi-branch CNN-feature fusion-XGBoost end-to-end hybrid model, as detailed below:

[0009] Input layer: Structured input, including spectral matrix, temperature field matrix, voltage curve matrix and 3D structured parameter vector;

[0010] Feature extraction layer: Three independent CNN branches operate in parallel, outputting three 128-dimensional feature vectors, which are concatenated and L2 normalized to form a 420-dimensional fused feature vector;

[0011] Step 3: During inference and evaluation, the multi-branch CNN-feature fusion-XGBoost end-to-end hybrid model is called. The fused features are input to obtain the loss rate prediction value. Combined with the SHAP value analysis, the core loss source is output to complete the quantitative attribution of low efficiency.

[0012] Step four: Match the model prediction results with the constructed loss threshold-core factors-operation and maintenance suggestion rule base, and push the operation and maintenance suggestions to the operation and maintenance platform. The operation and maintenance platform receives the decision instructions and executes the operation. At this point, the evaluation and optimization process is completed.

[0013] A photovoltaic power generation efficiency evaluation and optimization system based on the Internet of Things is provided. The system is used to execute the photovoltaic power generation efficiency evaluation and optimization method. The system adopts a model architecture of perception layer-edge layer-transmission layer-processing layer-decision layer-execution layer.

[0014] Preferably, the cloud server in the processing layer integrates data cleaning, feature engineering, and multi-branch CNN-feature fusion-XGBoost end-to-end hybrid model construction, and outputs loss rate, core loss sources, and efficiency evaluation results.

[0015] Preferably, the data cleaning uses a 3σ criterion-isolated forest algorithm joint identification mechanism to process abnormal data, and LSTM time series interpolation is used to complete samples with a missing rate of ≤5%.

[0016] Preferably, feature engineering includes mapping all numerical features using the Min-Max normalization method and reshaping spatial / sequence features.

[0017] Preferably, the decision-making level constructs a rule base for loss level, core factors, and operation and maintenance strategies, and dynamically generates optimization instructions based on real-time data.

[0018] Preferably, the execution layer is linked with the photovoltaic power plant operation and maintenance platform to perform optimization operations such as cleaning, heat dissipation, and component maintenance, and collects efficiency data after execution through IoT terminals.

[0019] The beneficial effects of this invention are as follows:

[0020] This invention constructs a "CNN-XGBoost" hybrid prediction architecture. Through a phased feature learning and multi-source information fusion strategy, it utilizes the spatial feature extraction advantage of CNN to accurately capture local correlation patterns of data such as spectra and temperature fields, and leverages the gradient boosting characteristics of XGBoost to fit the nonlinear coupling relationship between multiple factors. By using feature fusion, the advantages of the two types of models are complemented, thereby significantly improving the prediction accuracy and generalization ability of the model.

[0021] This invention solves the problems of traditional operation and maintenance relying on experience for decision-making, lagging efficiency assessment, and vague loss tracing. It achieves more accurate efficiency assessment and intelligent operation and maintenance decision-making through data-driven approaches. Field verification has shown that it can effectively improve the overall power generation efficiency of photovoltaic power plants. Detailed Implementation Example 1

[0022] The perception layer consists of component-level sensors, distributed fiber optic sensors, and environmental sensors;

[0023] (1) Component-level sensors: Each component is equipped with one BH1750 light sensor (with parameters: measurement range 0-65535lx, accuracy ±20%, sampling rate 10Hz) and one MS100 spectral sensor (with parameters: resolution 10nm, sampling rate 10Hz).

[0024] The BH1750 light sensor and MS100 spectral sensor are fixed to the top of the component frame (5cm from the edge, ensuring an occlusion rate of ≤0.1%) using an L-shaped aluminum bracket, ensuring that the acquisition angle is consistent with the component's light-receiving angle, and accurately reflecting the actual light resource reception of the component.

[0025] (2) Distributed fiber optic sensor: FOT-900 type fiber optic (i.e. distributed fiber optic sensor, model FOT-900, temperature measurement range -40℃~120℃, accuracy ±0.5℃, supports 2km long distance monitoring) is laid along the component array support to monitor the temperature field of the component array in full coverage.

[0026] Distributed optical fibers are laid along the bottom of the component array support and fixed with stainless steel clips every 1 meter to ensure that the optical fibers are parallel to the component array and that the temperature monitoring points are evenly distributed. Furthermore, heat shrink tubing and waterproof tape are used to seal the fiber splices to control the splice loss to ≤0.1dB and avoid signal attenuation that could reduce the accuracy of temperature measurement.

[0027] (3) Environmental sensors: Five monitoring points are set up according to the uniform grid principle of the 1:500 topology map of the power station. At the monitoring points, SHT30 temperature and humidity sensors (temperature measurement range -40℃~125℃, accuracy ±0.3℃; humidity measurement range 0-100%RH, accuracy ±2%RH; sampling rate 10Hz) + FS-10 wind speed sensors (measurement range 0-30m / s, accuracy ±0.2m / s, sampling rate 10Hz) are installed.

[0028] The temperature and humidity sensor and the wind speed sensor are installed at a height of 1.5m and avoid the shaded area of ​​the component to simulate the actual environmental conditions of the component's operation. This avoids the distortion of light, temperature and humidity data caused by shadows and is used to obtain macroscopic environmental parameters to provide an environmental benchmark for the correction of component operating data. Example 2

[0029] The transport layer consists of a LoRaWAN gateway, a 5G module, and a micro satellite terminal;

[0030] The LoRaWAN gateway (model LG300, deployed according to the standard of each gateway carrying 100-120 sensors) is fixed on a high-point bracket with the antenna facing the component array and connected to the data center switch via a network cable, and configured with a static IP and communication frequency band (470-510MHz).

[0031] The equipment room is equipped with a 5G module (model AG550Q) and a micro satellite terminal (model Kymeta u8, supporting Ku band, with a transmission rate of 1-10Mbps). Example 3

[0032] The edge layer consists of one standard edge terminal deployed and fixed in the power distribution box for every 10 component arrays. The edge terminal uses the STM32H7 model, which has a built-in LoRa module and is equipped with three types of core hardware interfaces: RS485 interface, fiber optic interface, and Ethernet interface. The RS485 interface connects to the BH1750 light sensor, MS100 spectrum sensor, SHT30 temperature and humidity sensor, and FS-10 wind speed sensor via the Modbus-RTU protocol. The fiber optic interface is directly connected to the FOT-900 distributed fiber optic sensor. The Ethernet interface is reserved for a wired connection channel with the data center switch.

[0033] The STM32H7 edge terminal establishes a connection with the nearest LG300 LoRaWAN gateway through its built-in LoRa module. It uploads compressed raw data to the gateway every 60 seconds. The uploaded data frame contains fields such as terminal ID, data type identifier, compressed package, and checksum.

[0034] The LG300 LoRaWAN gateway is configured with a dedicated port to receive compressed raw data frames uploaded by edge terminals. After receiving the data frames, it prioritizes sending data to the cloud MQTT Broker (message queue) through the AG550Q 5G module. When the 5G signal strength is ≤-90dBm, it automatically switches to satellite communication within 10 seconds. Example 4

[0035] The specific steps for constructing a photovoltaic module loss prediction model are as follows:

[0036] Step 1, data preprocessing, as follows:

[0037] By integrating data from the sensing layer (1,000 data points in the 300-1100nm spectral band, light intensity, 500 data points in the distributed temperature field, ambient temperature and humidity, and wind speed), power plant SCADA system operation and maintenance data (dust cleaning records and component replacement logs), and power grid data (voltage fluctuation curves), a raw dataset of 500,000 samples with 1,532 dimensions was formed.

[0038] The output power of the module array was measured using a FLUKE 438-II standard power analyzer (accuracy ±0.1%), and the theoretical power of the module array was calculated by combining the photovoltaic module IV characteristic curve. The difference is the loss tag value. For non-measured samples, loss tags were generated using the PSCAD / EMTDC simulation platform to ensure that the tag reliability is ≥99%.

[0039] A joint identification mechanism of the 3σ criterion and the Isolation Forest algorithm is adopted for anomaly data processing. Based on the assumption of normal distribution, the 3σ criterion is used to initially screen potential anomaly samples that exceed the mean ± 3 standard deviations. The Isolation Forest algorithm is then used to perform secondary verification on potential anomaly samples, and the anomaly data removal rate is finally controlled within 1.2%, ensuring the quality of the training dataset. This mechanism can effectively identify typical anomaly data such as sudden jump values ​​in spectral data and abnormally high temperature points in temperature fields, providing a basic guarantee for the stability of subsequent model training.

[0040] For samples with a missing value rate of ≤5%, LSTM time series interpolation is used for completion. Historical data from 5 time points before and after the target missing point are used as training samples to build an LSTM time series prediction model to learn the temporal dependency characteristics of the data, thereby achieving accurate prediction of missing values ​​and controlling the completion error within 2.3%.

[0041] The Min-Max normalization method is used for all numerical features (including light intensity, ambient temperature, wind speed and spectral band values, etc.) to map the data to the [0,1] interval. This operation can effectively eliminate the dimensional differences between different features (such as light intensity in lx and temperature in ℃), avoid interference from the differences in feature numerical ranges during model training, and ensure that each feature has a balanced weight during model training.

[0042] Step two, input feature classification and processing, as follows:

[0043] (1) Spatial / sequence features: including spectral data (1×1000-dimensional vector), distributed temperature field data (1×500-dimensional vector), and voltage fluctuation curves (1×N-dimensional vector, where N is the number of sampling points); these features have significant local correlations (such as intensity correlations between adjacent spectral bands and distribution correlations between adjacent measurement points in the temperature field), and are suitable for extracting their spatial pattern features through convolutional neural networks (CNN);

[0044] Spatial / Sequence Feature Reshaping: To meet the input dimensionality requirements of CNN models, spatial / sequence features in one-dimensional vector form are reshaped into two-dimensional matrix form.

[0045] Spectral data (1×1000): Divided into 100 groups according to the principle of 10 adjacent bands as a group, and reconstructed into a two-dimensional matrix of 10×100×1 to preserve the local correlation characteristics between bands;

[0046] Distributed temperature field data (1×500): Based on the spatial distribution characteristics of the component array, the measurement point matrix is ​​divided into 10 rows and 5 columns, and then reshaped into a 10×50×1 two-dimensional matrix to preserve the spatial location correlation of the temperature measurement points.

[0047] Voltage fluctuation curve (1×N): The timing dimension characteristics are directly preserved and reshaped into a 1×N×1 two-dimensional matrix to capture the timing variation trend of voltage.

[0048] (2) Structured features: These include environmental parameters (light intensity, ambient temperature and humidity, wind speed) and operation and maintenance data (dust cleaning records, component replacement logs). These features exist as independent numerical values ​​or categories, making them suitable for capturing the nonlinear relationship between the features and losses using the Extreme Gradient Boosting Tree (XGBoost) model.

[0049] Numerical structured features (light intensity, ambient temperature and humidity, wind speed): After Min-Max normalization, they are directly used as model input;

[0050] Categorical structured features (dust accumulation cleanup records: cleaned and uncleaned; component replacement logs: replaced and not replaced): are converted into binary vector form using one-hot encoding (e.g., cleaned is encoded as [1,0], uncleaned is encoded as [0,1]) to avoid the category order interfering with model training;

[0051] (3) A multi-branch CNN feature extraction-feature fusion-XGBoost prediction architecture is adopted to collaboratively learn spatial and nonlinear features and improve prediction performance;

[0052] Dedicated CNN branches were designed for the three types of features, with customized convolution kernels and pooling strategies, each outputting a 128-dimensional feature vector:

[0053] (3-1) CNN branch for spectral data: For the 10×100×1 input dimension of spectral data, a dedicated CNN branch is designed to extract spatial correlation features between bands:

[0054] Input dimensions: 10×100×1;

[0055] Network structures are as follows: Conv2D (3×3 convolutional kernels, 32 kernels, stride 1), MaxPool2D (2×2 pooling kernels, stride 2), Conv2D (3×3 convolutional kernels, 64 kernels, stride 1), MaxPool2D (2×2 pooling kernels, stride 2), Conv2D (3×3 convolutional kernels, 128 kernels, stride 1), and GlobalAvgPool2D.

[0056] Output features: 128-dimensional spectral feature vectors, which can effectively capture spatial pattern information such as local correlations and intensity gradients between spectral bands;

[0057] (3-2) CNN branch for temperature field data: For the 10×50×1 input dimension of the temperature field data, a dedicated CNN branch is designed to extract the spatial distribution features of the temperature measurement points:

[0058] Input dimensions: 10×50×1;

[0059] Network structure: Consistent with the CNN branches for spectral data, the branches are Conv2D(3×3, 32, 1), MaxPool2D(2×2, 2), Conv2D(3×3, 64, 1), MaxPool2D(2×2, 2), Conv2D(3×3, 128, 1) and GlobalAvgPool2D in sequence;

[0060] Output features: 128-dimensional temperature field feature vector, which can accurately capture information such as the spatial distribution pattern and gradient change characteristics of temperature measurement points;

[0061] (3-3) CNN branch for voltage fluctuation curve: For the 1×N×1 time series features of the voltage fluctuation curve, a one-dimensional CNN branch is designed to extract the time series correlation features:

[0062] Input dimension: 1×N×1;

[0063] Network structures are as follows: Conv2D (1×3 convolutional kernels, 32 kernels, stride 1), MaxPool2D (1×2 pooling kernels, stride 2), Conv2D (1×3 convolutional kernels, 64 kernels, stride 1), MaxPool2D (1×2 pooling kernels, stride 2), Conv2D (1×3 convolutional kernels, 128 kernels, stride 1), and GlobalAvgPool2D.

[0064] Output features: 128-dimensional voltage feature vector, which can effectively capture key features such as local trends and fluctuation frequency of voltage time series changes;

[0065] (4) Construct a three-level fusion process of branch feature concatenation, cross-type feature fusion, and feature normalization to form a unified model input feature set, as follows:

[0066] (4-1) CNN branch feature concatenation: The 128-dimensional feature vectors output by the three CNN branches of spectrum, temperature field and voltage curve are concatenated end to end according to dimension to obtain a 384-dimensional spatial feature vector, realizing the initial fusion of spatial / sequence features;

[0067] (4-2) Cross-type feature fusion: The 384-dimensional spatial feature vector and the 36-dimensional structured feature vector (environmental parameters and operation and maintenance data after standardization and encoding) are concatenated to form a 420-dimensional fused feature vector;

[0068] (4-3) Key optimization strategy: Before feature fusion, L2 normalization is performed on all feature vectors (the magnitude of the feature vector is normalized to 1) to eliminate the interference of scale differences between different types of features on the training of XGBoost model and accelerate the model convergence process.

[0069] (5) Using 420-dimensional features as input, Bayesian optimization is used to determine the optimal parameters, as follows:

[0070] Number of decision trees (n_estimators): 100;

[0071] Single tree depth (max_depth): 6;

[0072] Learning rate: 0.1;

[0073] Subsample ratio: 0.8;

[0074] Column sampling ratio (colsample_bytree): 0.8;

[0075] Regularization coefficients (reg_alpha=0.1, reg_lambda=0.1): used to suppress model overfitting;

[0076] Objective function: reg:squarederror (mean squared error loss function);

[0077] Using the root mean square error (RMSE) of the validation set as the monitoring metric, the training process is terminated when the RMSE of the validation set does not decrease significantly after 10 consecutive training rounds, effectively suppressing model overfitting.

[0078] The training set was randomly divided into 5 mutually exclusive subsets. Four subsets were selected as the training set and one subset as the validation set. The training and validation process was repeated 5 times. The final model performance was the average of the 5 results, which significantly improved the model's generalization ability.

[0079] Step 3, model training and optimization, is as follows:

[0080] The dataset is divided into: a training set (70%) for iterative learning and updating of model parameters; a validation set (20%) for monitoring overfitting during model training and providing a basis for hyperparameter tuning; and a test set (10%) for evaluating the final predictive performance of the model. The test set data was not involved in the model training and parameter tuning process to ensure the objectivity and reliability of the evaluation results.

[0081] Regression performance metrics include mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) for both training and validation sets. 2 );

[0082] Tuning strategy: Bayesian optimization of hyperparameters (tuning times ≤ 30), eliminating redundant features with importance < 0.001;

[0083] Step four: The trained photovoltaic module loss prediction model is a multi-branch CNN-feature fusion-XGBoost end-to-end hybrid model, with the following specific morphological parameters:

[0084] Input layer: Accepts four types of structured inputs, including a 10×100×1 spectral matrix, a 10×50×1 temperature field matrix, a 1×N×1 voltage curve matrix, and a 36-dimensional structured parameter vector (including environmental and operation and maintenance coding data).

[0085] Feature extraction layer: Three independent CNN branches operate in parallel, outputting three 128-dimensional feature vectors, which are concatenated and L2 normalized to form a 420-dimensional fused feature vector;

[0086] Prediction layer: XGBoost regressor (n_estimators=100, max_depth=6), outputs single-dimensional loss prediction (unit: kW).

[0087] Deployment format: Packaged in ONNX format using PyTorch+Scikit-learn framework, model size 128MB, supports CPU / GPU inference, single inference time ≤30ms;

[0088] Supporting modules: Integrates data preprocessing interfaces (outlier detection, missing value completion, standardization) and loss tracing interfaces, supporting end-to-end prediction and factor analysis. Example 5

[0089] The results of static validation using the test set are: loss prediction RMSE ≤ 3.8%, and coefficient of determination R0. 2 ≥0.95, loss tracing accuracy ≥97%, and mean absolute error (MAE) ≤2.5%;

[0090] The results of dynamic inference verification are as follows: response time ≤30ms, consistency with the measured loss value reaches 98.2%, and accuracy in extreme scenarios ≥95% (high temperature ≥45℃); the test environment consists of 5 component arrays tested for 10 days, covering 8 typical scenarios. Example 6

[0091] To address the structural characteristics of the XGBoost prediction layer and high-dimensional fusion features, the TreeExplainer interpreter is used to calculate the contribution of each loss factor. The process is as follows:

[0092] The 420-dimensional fused features are reverse-mapped into 12 categories of loss factors with clear physical meaning, including spectral features (mean intensity of band intensity, gradient variance), temperature factors (average temperature, maximum temperature, temperature gradient), voltage factors (fluctuation amplitude, mean deviation), environmental factors (light intensity, humidity, wind speed), and operation and maintenance factors (dust accumulation level, component aging time).

[0093] Initialize the TreeExplainer interpreter, set tree_limit=-1 and approximate=False, input the trained XGBoost model and validation set feature data, and output a 12×500 SHAP value matrix. In the matrix, rows correspond to loss factors, columns correspond to samples, positive and negative values ​​indicate whether the factors promote or inhibit loss, and absolute values ​​indicate the magnitude of contribution.

[0094] The average absolute SHAP value of each factor across all samples is calculated as a global contribution index to achieve comparability of the contributions of different factors, as detailed below: ;

[0095] in, MASH i For the first i The global contribution of each loss factor. SHAP i,j For the first i The factor affects the first j The SHAP value of each sample, N The number of samples;

[0096] Building a Loss Threshold-Factor Location Rule Base: Based on the core logic of loss threshold-based hierarchical control and precise matching of core factors, and combining SHAP value sample-level contribution analysis results with on-site operation and maintenance experience, a structured rule base is constructed to achieve closed-loop decision support: loss exceeding the limit early warning → automatic location of key factors → output of operation and maintenance suggestions. Specifically:

[0097] Based on the rated power of the components (taking 100kW as an example) and industry operation and maintenance standards, three levels of loss thresholds are defined, and the corresponding control levels for different thresholds are clearly defined, as follows:

[0098] Level 1 Threshold (Warning Level): Loss rate 5%-10% (corresponding to 5-10kW), requires close monitoring;

[0099] Level 2 Threshold (Intervention Level): Loss rate 10%-20% (corresponding to 10-20kW), requiring immediate operation and maintenance intervention;

[0100] Level 3 Threshold (Emergency Level): Loss rate ≥20% (corresponding to ≥20kW), emergency shutdown is required;

[0101] SHAP value clustering analysis was performed on samples within each threshold range to extract the three-element rule of loss level-core factor-contribution threshold. When the SHAP value of a certain factor accounts for ≥40% (single factor) or the SHAP value of a two-factor combination accounts for ≥60%, it is determined to be the core driving factor of that loss level.

[0102] The traceability results are integrated into the cloud management platform to develop a loss traceability dashboard, which displays the total loss value, the contribution ratio of each factor, and the location results of the core loss source in real time.

Claims

1. An Internet of Things-based photovoltaic power generation efficiency evaluation optimization method, characterized in that, Includes the following steps: Step 1: The IoT data acquisition layer is built to collect real-time spectral data, temperature field distribution data, voltage fluctuation curves, environmental parameters, and operation and maintenance records of photovoltaic modules, and transmit them to the processing layer through the edge layer and transmission layer; the spectral data consists of 1000 data points in the 300-1100nm band, and the temperature field distribution data consists of 500 data points collected based on distributed optical fiber sensors. Step 2: Construct a photovoltaic module loss prediction model. This model is a multi-branch CNN-feature fusion-XGBoost end-to-end hybrid model, as detailed below: Input layer: Structured input, including a reshaped spectral matrix, temperature field matrix, voltage curve matrix, and a 36-dimensional structured parameter vector, which includes environmental parameters and operation and maintenance record data; The spectral data is reshaped into a 10×100×1 two-dimensional matrix by grouping 10 adjacent bands. The temperature field distribution data is divided into 10 rows and 50 columns according to the spatial distribution of the component array and reshaped into a 10×50×1 two-dimensional matrix. The voltage fluctuation curve retains the time dimension and is reshaped into a 1×N×1 two-dimensional matrix, where N is the number of sampling points. Feature extraction layer: Three independent CNN branches perform parallel operations on spectral data, temperature field distribution data, and voltage fluctuation curves respectively, outputting three 128-dimensional feature vectors. These are concatenated to form a 384-dimensional spatial feature vector, which is then concatenated with the 36-dimensional structured parameter vector across types and normalized by L2 to form a 420-dimensional fused feature vector. Step 3: During inference and evaluation, the multi-branch CNN-feature fusion-XGBoost end-to-end hybrid model is called. The fused feature vector is input to obtain the loss rate prediction value. The TreeExplainer interpreter is used to calculate the SHAP value of each loss factor. The 420-dimensional fused feature vector is reverse-mapped into loss factors with clear physical meaning. The core loss source is analyzed and output in combination with the SHAP value to complete the quantitative attribution of low efficiency. Step four: Match the model prediction results with the constructed loss threshold-core factor-operation and maintenance suggestion rule base. The rule base is divided into three levels of loss thresholds based on the rated power of the components, and includes a three-element rule of loss level-core factor-contribution threshold. When the SHAP value of a certain factor accounts for ≥40% or the combined SHAP value of two factors accounts for ≥60%, it is determined to be a core driving factor, and operation and maintenance suggestions are pushed to the operation and maintenance platform. The operation and maintenance platform receives the decision instructions and executes the operation. At this point, the evaluation and optimization process is completed.

2. A photovoltaic power generation efficiency evaluation and optimization system based on the Internet of Things (IoT), the system being used to execute the photovoltaic power generation efficiency evaluation and optimization method based on the IoT as described in claim 1, characterized in that, The system adopts a model architecture of perception layer-edge layer-transmission layer-processing layer-decision layer-execution layer.

3. The photovoltaic power generation efficiency evaluation and optimization system based on the Internet of Things according to claim 2, characterized in that, The sensing layer consists of component-level sensors, distributed fiber optic sensors, and environmental sensors, including temperature and humidity sensors and wind speed sensors.

4. The photovoltaic power generation efficiency evaluation and optimization system based on the Internet of Things according to claim 2, characterized in that, The edge terminal of the edge layer has a built-in LoRa module and is configured with three types of core hardware interfaces: RS485 interface, fiber optic interface, and Ethernet interface.

5. The photovoltaic power generation efficiency evaluation and optimization system based on the Internet of Things according to claim 2, characterized in that, The transport layer consists of a LoRaWAN gateway, a 5G module, and a micro satellite terminal.

6. The photovoltaic power generation efficiency evaluation and optimization system based on the Internet of Things according to claim 2, characterized in that, The cloud server in the processing layer integrates data cleaning, feature engineering, and multi-branch CNN-feature fusion-XGBoost end-to-end hybrid model construction, and outputs loss rate, core loss sources, and efficiency evaluation results.

7. The photovoltaic power generation efficiency evaluation and optimization system based on the Internet of Things according to claim 6, characterized in that, Data cleaning employs a 3σ criterion-isolated forest algorithm joint identification mechanism for outlier data processing, and LSTM time series interpolation is used to complete samples with a missing rate ≤5%.

8. The photovoltaic power generation efficiency evaluation and optimization system based on the Internet of Things according to claim 6, characterized in that, Feature engineering includes mapping all numerical features using the Min-Max normalization method and reshaping spatial / sequence features.

9. The photovoltaic power generation efficiency evaluation and optimization system based on the Internet of Things according to claim 2, characterized in that, The decision-making level constructs a rule base for loss thresholds, core factors, and operation and maintenance suggestions, and dynamically generates optimization instructions based on real-time data.

10. The photovoltaic power generation efficiency evaluation and optimization system based on the Internet of Things according to claim 2, characterized in that, The execution layer works in conjunction with the photovoltaic power plant operation and maintenance platform to perform cleaning, heat dissipation, and component maintenance and optimization operations, and collects efficiency data after execution through IoT terminals.