Distributed photovoltaic power generation prediction scheduling method and system

Abstract

The invention discloses a distributed photovoltaic power generation prediction scheduling method and system, and the method comprises the steps: storing and indexing multi-source meteorological data and photovoltaic system historical operation data through a link array tree data structure, constructing a time sequence feature extraction model, a spatial feature extraction model and a parameter feature extraction model, and carrying out the fusion through an integrated learning method, and meanwhile, a photovoltaic module aging analysis model is introduced to correct a prediction result, short-term, medium-term and long-term multi-time-scale power generation capacity prediction is realized, and a power grid interaction scheme and an energy storage control strategy are generated. According to the method, the problems that in the prior art, a single model is difficult to process different time scale prediction tasks at the same time, the data retrieval efficiency is low, and aging attenuation of the photovoltaic module is neglected are solved, and the prediction accuracy and the scheduling efficiency of the distributed photovoltaic system are improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation technology, and in particular to a method and system for predictive scheduling of distributed photovoltaic power generation. Background Technology

[0002] As renewable energy continues to increase its share in the global energy mix, distributed photovoltaic (PV) power generation, with its flexibility, environmental friendliness, and economic efficiency, has become an important direction for clean energy development. Distributed PV power generation systems are typically installed on the user side, directly providing electricity to users. Excess electricity can be sold to the grid, featuring local consumption and decentralized layout.

[0003] Currently, common photovoltaic (PV) power generation forecasting technologies mainly include physical model-based methods and data-driven methods. Physical model-based methods establish power generation forecasting models based on the physical characteristics of PV cells and meteorological factors such as irradiance and temperature. Data-driven methods, on the other hand, collect historical meteorological and power generation data and use statistical learning or machine learning methods to construct forecasting models. For example, some systems use support vector machines (SVM) or simple neural network models to process irradiance and temperature data to predict short-term PV output power.

[0004] In recent years, deep learning-based photovoltaic power generation forecasting methods have emerged. By constructing deep neural network models to learn from historical power generation and meteorological data, these methods can improve forecast accuracy to some extent. These methods typically employ a single deep learning model, such as a Long Short-Term Memory (LSTM) network or a Convolutional Neural Network (CNN), to analyze and process time-series data and output predicted power generation values ​​for a future period.

[0005] However, existing technologies have the following significant shortcomings: First, a single model struggles to simultaneously handle forecasting tasks at different time scales (short-term, medium-term, and long-term), limiting forecast accuracy and applicability. Second, existing systems typically use traditional databases or simple data structures to store and retrieve historical data, resulting in low retrieval efficiency when faced with massive amounts of multi-source meteorological and power generation data, making it difficult to meet real-time scheduling requirements. Furthermore, existing systems lack consideration for the aging and degradation of photovoltaic modules, failing to accurately reflect performance changes over time and impacting forecast accuracy. These problems severely restrict the efficient scheduling and widespread application of distributed photovoltaic systems. Summary of the Invention

[0006] The purpose of this invention is to provide a distributed photovoltaic power generation prediction and scheduling method and system, which aims to solve the problems in the prior art, such as the difficulty of a single model to handle prediction tasks at different time scales simultaneously, low data retrieval efficiency, and neglect of photovoltaic module aging and degradation.

[0007] To achieve the above objectives, this invention provides a distributed photovoltaic (PV) power generation prediction and scheduling method, comprising: acquiring multi-source meteorological monitoring data and historical operating data of the PV system; performing data cleaning and standardization on the multi-source meteorological monitoring data and the historical operating data of the PV system to obtain standardized training data; constructing a linked array tree data structure of a multi-dimensional feature space based on the standardized training data; storing and indexing the standardized training data to generate a data access interface; acquiring historical data using the data access interface; constructing a time-series feature extraction model, a spatial feature extraction model, and a parametric feature extraction model based on the historical data; and using ensemble learning... The model integrates the outputs of the time-series feature extraction model, the spatial feature extraction model, and the parametric feature extraction model to obtain preliminary prediction results. Based on the historical operating data of the photovoltaic system, a photovoltaic module aging analysis model is established. A performance correction factor is calculated based on the photovoltaic module aging analysis model to correct the preliminary prediction results, resulting in corrected prediction results. Based on the corrected prediction results, short-term, medium-term, and long-term power generation predictions are performed respectively to generate multi-time-scale prediction results. Based on the multi-time-scale prediction results and combined with electricity market data, a grid interaction scheme and energy storage control strategy are generated to form a photovoltaic power generation dispatch strategy.

[0008] Furthermore, the acquisition of multi-source meteorological monitoring data and historical operating data of the photovoltaic system, followed by data cleaning and standardization of the multi-source meteorological monitoring data and the historical operating data of the photovoltaic system to obtain standardized training data, includes: acquiring irradiance, temperature, humidity, cloud cover, and wind speed data from multiple meteorological monitoring stations, and generating unified time-series meteorological data through data synchronization processing; collecting power generation data and equipment operating parameters from the photovoltaic system, performing outlier detection and interpolation processing to generate cleaned historical data; and standardizing the unified time-series meteorological data and the cleaned historical data to generate the standardized training data.

[0009] Furthermore, based on the standardized training data, a linked array tree data structure in a multi-dimensional feature space is constructed. The standardized training data is stored and indexed, and a data access interface is generated. This includes: partitioning the standardized training data into a feature space, constructing a tree-structured storage framework, and generating multi-level nodes; storing the standardized training data into the corresponding tree-structured nodes, establishing links between nodes to form a linked array tree; and designing a data update algorithm to dynamically maintain the linked array tree and keep the data structure balanced.

[0010] Further, the step of acquiring historical data using the data access interface, constructing a time-series feature extraction model, a spatial feature extraction model, and a parametric feature extraction model based on the historical data, and fusing the outputs of the time-series feature extraction model, the spatial feature extraction model, and the parametric feature extraction model through an ensemble learning model to obtain preliminary prediction results includes: training the time-series feature extraction model based on the historical data to generate time-series feature extraction results; processing the spatial distribution characteristics of meteorological data using the spatial feature extraction model to obtain spatial feature extraction results; constructing a parametric feature extraction model to process photovoltaic module parameters to generate module characteristic prediction results; and fusing the time-series feature extraction results, spatial feature extraction results, and module characteristic prediction results using an ensemble learning model to obtain the preliminary prediction results.

[0011] Furthermore, the step of establishing a photovoltaic module aging analysis model based on the historical operating data of the photovoltaic system, calculating a performance correction factor based on the photovoltaic module aging analysis model, and correcting the preliminary prediction results to obtain the corrected prediction results includes: analyzing historical power generation data and operating environment data of the photovoltaic module based on the historical operating data of the photovoltaic system to generate performance degradation analysis results; analyzing and processing environmental factors to obtain environmental impact analysis results; and fusing the performance degradation analysis results and the environmental impact analysis results to generate the photovoltaic module aging analysis model.

[0012] Furthermore, based on the revised prediction results, short-term, medium-term, and long-term power generation predictions are performed respectively to generate multi-timescale prediction results, including: using a short-term prediction model to process real-time meteorological data to generate short-term power generation prediction values; analyzing short-term meteorological forecast data based on a medium-term prediction model to generate medium-term power generation prediction values; and using a long-term prediction model to process seasonal variation data to generate long-term power generation prediction values.

[0013] Furthermore, based on the multi-timescale prediction results and combined with electricity market data, a grid interaction scheme and energy storage control strategy are generated to form a photovoltaic power generation dispatch strategy, including: using historical electricity consumption data to predict the electricity load curve and obtain electricity demand prediction results; calculating the grid interaction period according to the electricity price policy and the electricity demand prediction results to generate a power generation consumption scheme; and optimizing the energy storage system charging and discharging plan based on the power generation consumption scheme to form the photovoltaic power generation dispatch strategy.

[0014] Furthermore, the method also includes: collecting actual operating data of the photovoltaic system, calculating the deviation between the predicted result and the actual value, and generating prediction deviation data; based on the prediction deviation data, updating the prediction model parameters using an online learning method to improve prediction accuracy.

[0015] Furthermore, the method also includes: monitoring the operating status of the photovoltaic system, detecting abnormal operating conditions, and generating an abnormal status report; dynamically adjusting the photovoltaic power generation scheduling strategy based on the abnormal status report to ensure the safe operation of the system.

[0016] This invention also provides a distributed photovoltaic power generation prediction and scheduling system, comprising: a data acquisition and preprocessing module, used to acquire multi-source meteorological monitoring data and historical operation data of the photovoltaic system, and to perform data cleaning and standardization processing on the multi-source meteorological monitoring data and the historical operation data of the photovoltaic system to obtain standardized training data; a data structure construction module, used to construct a linked array tree data structure of a multi-dimensional feature space based on the standardized training data, to store and index the standardized training data, and to generate a data access interface; and a model construction and fusion module, used to acquire historical data using the data access interface, to construct a time-series feature extraction model, a spatial feature extraction model, and a parametric feature extraction model based on the historical data, and to perform ensemble learning. The model integrates the outputs of the time-series feature extraction model, the spatial feature extraction model, and the parametric feature extraction model to obtain preliminary prediction results. The aging correction module establishes a photovoltaic module aging analysis model based on the historical operating data of the photovoltaic system, calculates a performance correction factor based on the photovoltaic module aging analysis model, and corrects the preliminary prediction results to obtain corrected prediction results. The prediction execution module performs short-term, medium-term, and long-term power generation predictions based on the corrected prediction results, generating multi-time-scale prediction results. The scheduling strategy module, based on the multi-time-scale prediction results and combined with electricity market data, generates a grid interaction scheme and an energy storage control strategy to form a photovoltaic power generation scheduling strategy.

[0017] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0018] 1. By constructing a linked array tree data structure, efficient storage and rapid retrieval of massive multi-source meteorological and power generation data are achieved, significantly improving data access efficiency and providing strong support for real-time scheduling.

[0019] 2. By adopting multi-model fusion technology, the temporal feature extraction model, spatial feature extraction model and parametric feature extraction model are organically combined to make full use of information from various types of data, thereby improving the accuracy and robustness of prediction.

[0020] 3. A photovoltaic module aging analysis model is introduced to consider the performance degradation of the module over time. The prediction results are corrected by a dynamic correction factor, which solves the problem that the existing technology ignores the impact of module aging.

[0021] 4. It enables multi-timescale forecasting of short, medium and long term, providing comprehensive guidance for the daily operation of photovoltaic systems, weekly planning and monthly energy planning, and enhancing the system's adaptability and foresight.

[0022] 5. A closed-loop feedback optimization mechanism was constructed. By collecting operational data in real time, analyzing prediction deviations, and updating model parameters, the system can achieve self-learning and continuous optimization, thereby improving prediction accuracy and scheduling efficiency. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a distributed photovoltaic power generation prediction and scheduling method according to the present invention;

[0025] Figure 2 This is a schematic diagram of the linked array tree data structure of the present invention;

[0026] Figure 3 This is a diagram of the multi-artificial intelligence model fusion architecture of the present invention;

[0027] Figure 4 This is a structural block diagram of the distributed photovoltaic power generation prediction and scheduling system of the present invention. Detailed Implementation

[0028] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0029] like Figure 1 As shown, the present invention provides a distributed photovoltaic power generation prediction and scheduling method, comprising the following steps:

[0030] Step S1: Obtain multi-source meteorological monitoring data and historical operation data of photovoltaic system, perform data cleaning and standardization processing on the multi-source meteorological monitoring data and the historical operation data of photovoltaic system to obtain standardized training data.

[0031] In this embodiment, step S1 specifically includes: acquiring irradiance, temperature, humidity, cloud cover, and wind speed data from multiple meteorological monitoring stations, and generating unified time-series meteorological data through data synchronization processing; collecting power generation data and equipment operating parameters from the photovoltaic system, performing outlier detection and interpolation processing to generate cleaned historical data; and standardizing the unified time-series meteorological data and the cleaned historical data to generate standardized training data.

[0032] Specifically, during the multi-source meteorological data acquisition process, real-time and historical meteorological data, including solar irradiance (W / m²), are obtained from multiple meteorological monitoring stations and third-party meteorological service interfaces. 2 Parameters such as ambient temperature (°C), relative humidity (%), cloud cover (%), wind speed (m / s), wind direction (°), and precipitation (mm) are collected, and meteorological forecast data are acquired to establish a complete meteorological database.

[0033] For historical data collection of photovoltaic systems, historical power generation data (kWh), DC side voltage (V), current (A), AC side active power (kW), reactive power (kVar), module temperature (°C) and other operating parameters are collected through the photovoltaic system monitoring interface. At the same time, static information such as photovoltaic module specifications, installation angle, and geographical location are also collected.

[0034] In the process of data anomaly detection and cleaning, an outlier detection algorithm based on statistical analysis is designed to identify and process missing values, outliers and noise in the data. Anomalous data is repaired or removed by interpolation algorithms, filtering algorithms and cluster analysis to generate a clean dataset.

[0035] For data standardization and feature extraction, time synchronization and spatial interpolation are performed on the cleaned multi-source data to unify data from different sources and with different sampling frequencies onto a standard time series. Numerical features are normalized using methods such as Min-Max standardization and Z-score standardization. At the same time, temporal features (such as time of day, seasons of the year) and spatial features (such as geolocation coding) are extracted.

[0036] Finally, the processed dataset is divided into training, validation, and test sets according to time sequence. The training set is used for model training, the validation set is used for model parameter tuning, and the test set is used for model performance evaluation, ensuring the temporal continuity and representativeness of the dataset.

[0037] Step S2: Based on the standardized training data, construct a linked array tree data structure for the multidimensional feature space, store and index the standardized training data, and generate a data access interface.

[0038] In this embodiment, step S2 specifically includes: dividing the standardized training data into feature spaces, constructing a tree-structured storage framework, and generating multi-level nodes; storing the standardized training data into the corresponding tree-structured nodes, establishing link relationships between nodes, and forming a linked array tree; designing a data update algorithm to dynamically maintain the linked array tree and keep the data structure balanced.

[0039] like Figure 2As shown, when designing the linked array tree data structure, the system designs a special multidimensional index data structure called a linked array tree (LAT). Each node contains an array and multiple link pointers. The array stores data records with similar characteristics, and the link pointers connect to nodes with different characteristics.

[0040] In the process of multidimensional feature space partitioning, the feature space of meteorological and power generation data is divided into multiple dimensions (such as time dimension, meteorological parameter dimension, geographical location dimension, etc.). Each dimension is divided according to specific rules to form a gridded representation of the feature space.

[0041] For tree structure construction, the skeleton of the tree structure is constructed based on the feature space partitioning results. The top-level node represents the coarsest-grained data partitioning, the bottom-level node represents the finest-grained data partitioning, and adjacent levels are connected by link pointers.

[0042] In the data recording, storage, and linking stage, standardized historical data records are stored in the corresponding tree node array, and links between nodes are established so that related data can be quickly accessed through links without traversing the entire data structure.

[0043] For the dynamic maintenance mechanism, an incremental update algorithm is designed to support real-time access of new data and periodic cleanup of old data, maintain the balance and retrieval efficiency of the linked array tree, and implement data version control and historical backtracking functions.

[0044] The core design concept of Linked Array Trees (LATs) is to organize similar data records in a multi-dimensional feature space together and establish relationships between nodes through link pointers, enabling efficient data retrieval and updates. In its implementation, each LAT node consists of three parts: node header information, a data record array, and a set of link pointers. The node header information includes metadata such as the node's dimensional feature value range, total number of records, creation timestamp, and last update timestamp. The data record array uses a contiguous memory storage structure to store original data records with similar characteristics. The set of link pointers contains references to related nodes, which are established based on different feature relationships, such as temporal adjacency, spatial proximity, or parameter similarity.

[0045] LAT uses three levels of nodes: root nodes, dimension nodes, and leaf nodes. The root node is the entry point of the entire tree, dimension nodes correspond to one or more dimensions of the feature space, and leaf nodes directly store the original data records. The feature space partitioning employs an adaptive grid strategy, dynamically adjusting the grid size based on the data distribution density. For high-density areas (such as periods of drastic weather change), the system uses a finer-grained grid; for low-density areas (such as nighttime periods without power generation), a coarser-grained grid is used, balancing retrieval efficiency and storage overhead.

[0046] In terms of data organization, meteorological and power generation data are categorized and stored according to multiple dimensions. The main dimensions include time (year, month, day, hour), spatial dimensions (longitude, latitude, altitude), meteorological parameter dimensions (irradiance range, temperature range, humidity range, etc.), and system parameter dimensions (component type, installation angle, etc.). For the time dimension, a hierarchical time index is used, creating four-level index nodes for year, month, day, and hour, and establishing forward and backward links between nodes at the same time scale to support fast time-series queries. For the spatial dimension, a quadtree structure is used to recursively divide the two-dimensional geographic space until a preset minimum grid size or a threshold for the number of data points within a grid is reached, achieving efficient management of spatial data.

[0047] Data retrieval is one of the core functions of LAT. The system implements various query interfaces, including point query (retrieving data at a specific time point or location), range query (retrieving data within a time period or region), and nearest neighbor query (retrieving data most similar to given conditions). Point query first determines the path to the target node based on the query conditions, starting from the root node and proceeding down level by level according to the dimension partitioning rules until a leaf node containing the target data is reached. The average time complexity is O(log n), where n is the total amount of data. Range query requires identifying all nodes intersecting with the query range. The system uses a pruning strategy to quickly eliminate non-intersecting branches, reducing the number of nodes traversed. Nearest neighbor query uses a priority queue method, exploring neighboring nodes step by step according to distance priority to find the k closest data points.

[0048] Data update operations include insertion, deletion, and modification. When inserting new data, the leaf node position where the data should be stored is first determined. If the target node's data record array is not full, the new record is directly added to the end of the array; if the array is full, a node split operation is triggered, dividing the node's data into two parts according to the selected dimension, creating a new node, and updating the relevant links. Data deletion operations mark the target record as deleted and physically remove it when appropriate. Regular maintenance checks the data density of nodes, performing merge operations on overly sparse nodes and rebalancing operations on overly dense nodes to maintain overall structural balance and retrieval efficiency.

[0049] To support efficient processing of time-series data, the system implements a sliding window mechanism in LAT. Each time dimension node maintains a fixed-size circular buffer to store recently accessed data records. When performing continuous time-series queries (such as retrieving hourly power generation data for the past 7 days), the system can directly retrieve most of the data from the buffer, significantly improving query performance. Test results show that compared to traditional relational databases, LAT improves speed by 15-20 times in large-scale meteorological and power generation data queries; and compared to NoSQL databases, it improves speed by 5-8 times in complex multi-condition queries.

[0050] In terms of data reliability, LAT implements a two-tier persistence strategy. The in-memory data structures periodically serialize updated portions to disk, forming an incremental log; simultaneously, the system periodically performs full snapshot backups, recording the complete state of the data structures. When the system restarts or recovers from a failure, the most recent full snapshot is loaded first, and then the update operations from the incremental log are applied, ensuring data consistency and integrity. This design minimizes the impact of I / O operations on system performance while ensuring data security.

[0051] Step S3: Obtain historical data using the data access interface, construct a time-series feature extraction model, a spatial feature extraction model, and a parametric feature extraction model based on the historical data, and fuse the outputs of the time-series feature extraction model, the spatial feature extraction model, and the parametric feature extraction model through an ensemble learning model to obtain preliminary prediction results.

[0052] In this embodiment, the present invention first uses the data access interface constructed in step S2 to efficiently retrieve historical data. Compared with traditional databases, the linked array tree structure can quickly locate and obtain the required historical meteorological data and power generation records with near constant time complexity, greatly improving data acquisition efficiency. Based on prediction needs, the system selectively extracts historical data for different time periods and weather types, providing targeted training samples for subsequent model construction.

[0053] Based on the acquired historical data, a temporal feature extraction model is constructed. This model mainly consists of three deep learning networks: a Long Short-Term Memory (LSTM) network responsible for capturing long-term temporal dependencies, capable of "remembering" important patterns in distant historical data; a Temporal Convolutional Network (TCN) extracting features across different time spans through multi-scale convolutional operations; and a Transformer model utilizing a self-attention mechanism to handle long-distance associations in the sequence. These three networks process patterns at different time scales, and then their respective feature outputs are weighted and combined to form a complete temporal feature representation.

[0054] Meanwhile, the system constructs a spatial feature extraction model to process the spatial distribution features of meteorological data. This model employs a multi-layer convolutional neural network (CNN) structure, taking the two-dimensional spatial distribution of meteorological parameters as input and extracting local and global spatial features through multi-layer convolution and pooling operations. Specifically, the model assigns higher weights to meteorological elements that significantly affect photovoltaic power generation, such as cloud cover distribution and pressure gradient, enhancing their contribution to the feature extraction process. The model ultimately outputs a meteorological spatial distribution feature vector, providing spatial dimensional information for prediction.

[0055] The system also constructs a parameter feature extraction model to process photovoltaic module-related parameters. This model employs a deep regression network structure, receiving photovoltaic module specifications (such as module type, rated power, temperature coefficient, etc.), installation conditions (such as tilt angle, orientation, shading conditions, etc.), and historical performance data as input. Through nonlinear transformation of a multi-layer fully connected network, the model learns the complex relationship between module parameters and power generation efficiency, ultimately outputting the expected power generation efficiency coefficient under given meteorological conditions.

[0056] For different forecasting tasks, the system performs specialized training on the aforementioned basic models. Specifically, based on different combinations of forecast timescales (short-term, medium-term, and long-term) and weather types (sunny, cloudy, rainy, etc.), the system trains multiple sets of specialized models. For example, the short-term forecasting model for sunny days mainly focuses on minute changes in irradiance, while the long-term forecasting model for cloudy and rainy days pays more attention to the evolution of weather systems. Each set of specialized models is optimized under specific conditions to achieve the best forecasting performance.

[0057] Finally, an ensemble learning framework is constructed to fuse the outputs of multiple models. This framework employs a stacking strategy, using the predictions of each base model and specialized model as features, which are then input into a meta-learner constructed using Gradient Boosting Tree (GBDT). The meta-learner learns the reliability of each model under different conditions through cross-validation, automatically assigns optimal model weights, and generates the final fused prediction result. This ensemble approach fully leverages the strengths of each model, significantly improving the accuracy and robustness of predictions.

[0058] In this embodiment, a three-layer stacked Long Short-Term Memory (LSTM) network was designed and constructed for the specific implementation of the time-series feature extraction model. The first layer of this LSTM network contains 128 hidden units, the second layer contains 64 hidden units, and the third layer contains 32 hidden units. Reducing the number of hidden units layer by layer helps the model gradually extract higher-level abstract features. To prevent overfitting, a dropout layer is added after each LSTM layer, with a dropout rate set to 0.3. The model input consists of time-series data covering the past 168 hours (7 days), including meteorological parameters (irradiance, temperature, humidity, etc.) and historical power generation, forming a multi-dimensional feature vector at each time point. The LSTM model effectively captures dependencies in long-term series through a gating mechanism, particularly demonstrating strong memory capabilities for seasonal and diurnal variation patterns.

[0059] Simultaneously, a Temporal Convolutional Network (TCN) was constructed to extract multi-scale temporal features. The TCN model employs four causal convolutional layers with a kernel size of 5 and dilation rates of 1, 2, 4, and 8, respectively. The exponentially increasing dilation rate expands the receptive field, enabling the model to extract features at different time scales. ReLU activation and batch normalization layers are used after each convolutional layer to accelerate training convergence. Residual connections are also introduced to effectively alleviate the gradient vanishing problem in deep networks and improve the model's learning ability. The TCN model is particularly adept at capturing short-term local change patterns, such as rapid changes in irradiance caused by cloud movement.

[0060] The Transformer model is used to handle long-range dependencies in sequences. The implemented Transformer encoder contains four self-attention layers, each with eight attention heads and a dimension of 64. Position encoding uses sine and cosine functions, enabling the model to perceive the relative position of data within the time series. The feedforward neural network consists of two fully connected layers, with the intermediate layer having a dimension of 256 and using the ReLU activation function. Through its self-attention mechanism, the Transformer model can adaptively learn dependencies between any positions in a sequence, unaffected by distance, making it particularly suitable for capturing non-periodic long-range patterns.

[0061] During model training, a two-stage training strategy was employed. In the first stage, the three models were pre-trained independently, using 80% of the historical dataset as the training set and 20% as the validation set. The LSTM and TCN models used mean squared error (MSE) as the loss function, while the Transformer model used mean squared logarithmic error (MSLE) to better handle scaling variations in predicted values. The optimizer was Adam, with an initial learning rate of 0.001 and a learning rate decay strategy, reducing the learning rate to 0.5 every 30 epochs. The batch size was set to 64, and training continued for at least 200 epochs or until the validation loss no longer decreased.

[0062] In the second stage, a feature fusion network was designed to connect the outputs of the three pre-trained models. Feature fusion was performed through a two-layer fully connected network (100 neurons in the first layer and 50 neurons in the second layer), and finally, an output layer generated the final predicted value. In this stage, the parameters of the pre-trained models were frozen, and only the fusion network part was trained. The Huber loss function was used, which combines the advantages of MSE and MAE and is more robust to outliers. The model employed an early stopping strategy, stopping training when the validation loss no longer decreased for 10 consecutive epochs.

[0063] The model evaluation employs a multi-indicator comprehensive evaluation method, including root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 On the test dataset, the fused temporal feature extraction model significantly improved across all metrics compared to the single model, with RMSE reduced by 17.3%, MAPE reduced by 15.8%, and R... 2 This represents a 12.5% ​​improvement. To adapt to different weather types, specialized models were also trained for sunny, cloudy, and rainy days, further enhancing prediction accuracy.

[0064] The output of the time-series feature extraction model is the predicted power generation time series for each future time point (short-term prediction: the next 24 hours; medium-term prediction: the next 7 days; long-term prediction: the next 1-3 months), including predicted values ​​and confidence intervals. These results will serve as important inputs to the ensemble learning model in step S3, fusing with the results of the spatial feature extraction model and the parametric feature extraction model to form more accurate prediction results.

[0065] In the specific implementation of the spatial feature extraction model, a deep convolutional neural network structure specifically designed to handle the spatial distribution characteristics of meteorological data is employed. This model uses a 50km × 50km gridded area as the basic unit, mapping meteorological data within the area to a 128 × 128 two-dimensional tensor. Each channel corresponds to a meteorological parameter, including irradiance, temperature, humidity, cloud cover, wind speed, and air pressure. To handle interpolation issues at different geographical locations, the system uses a bilinear interpolation method to ensure the continuity and smoothness of the data distribution. The input to the convolutional neural network is a multi-channel two-dimensional tensor, with each channel representing the spatial distribution of a meteorological parameter.

[0066] The backbone network of this spatial feature extraction model employs an improved VGG architecture, containing five convolutional blocks. The first convolutional block consists of two convolutional layers, each using 64 3×3 convolutional kernels with a stride of 1 and padding of 1, maintaining the output feature map size unchanged. The second convolutional block also contains two convolutional layers, but uses 128 convolutional kernels. The third, fourth, and fifth convolutional blocks each contain three convolutional layers, using 256, 512, and 512 convolutional kernels, respectively. Each convolutional block is followed by a max-pooling layer with a 2×2 kernel size and a stride of 2, halving the feature map size. A ReLU activation function is used after all convolutional layers to increase the model's non-linear expressiveness, and a batch normalization layer is added after each convolutional block to accelerate training convergence.

[0067] To better capture the spatial distribution characteristics of meteorological elements that significantly impact photovoltaic power generation, such as cloud cover and irradiance, the model introduces an attention mechanism after the third convolutional block. Specifically, this is implemented as a spatial attention module, which generates a spatial attention map using a 1×1 convolutional layer and a sigmoid activation function. This module weights the feature map, enabling the model to automatically focus on features in important regions. This mechanism significantly improves the model's ability to perceive key influencing factors of photovoltaic power generation, such as cloud movement and shading effects.

[0068] At the top layer of the convolutional network, a global average pooling layer is added to compress the two-dimensional feature map into a one-dimensional feature vector. Then, the feature is transformed through two fully connected layers (512 neurons in the first layer and 256 neurons in the second layer), finally outputting a 128-dimensional spatial feature vector. This feature vector encodes the spatial distribution pattern of meteorological parameters and their interrelationships, providing important spatial dimensional information for subsequent power generation prediction.

[0069] During the model training phase, the system selected the stochastic gradient descent optimizer, with an initial learning rate of 0.01, a weight decay parameter of 0.0005, and a momentum parameter of 0.9. A learning rate scheduling strategy was used, reducing the learning rate to 0.1 every 50 epochs. The training batch size was 32, and the training process lasted for 300 epochs. To enhance the model's generalization ability, data augmentation techniques were employed during training, including random rotation (±15 degrees), horizontal flipping, and random pruning, to simulate different changes in the meteorological system. The loss function was weighted mean squared error, assigning higher weights to key parameters such as irradiance and cloud cover, guiding the model to focus more on these meteorological elements that significantly impact photovoltaic power generation.

[0070] To adapt to the meteorological characteristics of different geographical regions, the system also incorporates a transfer learning approach. First, the model is pre-trained on a large-scale meteorological dataset, then fine-tuned for data from specific regions. This allows the model to quickly adapt to the meteorological characteristics of new areas, significantly improving its applicability and deployment efficiency. Test results show that, compared to a baseline model without spatial features, the introduction of spatial feature extraction improves prediction accuracy by 23.5% under cloudy weather conditions and by 19.7% under rapidly changing weather conditions.

[0071] The parameter feature extraction model is specifically designed to process photovoltaic module specifications, installation conditions, and historical performance data, extracting the complex relationships between these parameters and power generation efficiency. The model input includes three types of features: static parameter features, installation condition features, and performance history features. Static parameter features include inherent module parameters such as module type (monocrystalline silicon, polycrystalline silicon, thin-film, etc.), rated power, temperature coefficient, and fill factor; installation condition features include installation environment parameters such as installation tilt angle, orientation angle, shading conditions, and dust accumulation; and performance history features include records of efficiency performance under different past conditions, such as actual power generation efficiency data under different temperatures and irradiance levels.

[0072] The system first preprocesses the input features. Categorical features (such as component type) are one-hot encoded; numerical features (such as rated power and temperature coefficient) are normalized to fall within the 0-1 range; installation angle features are transformed using trigonometric functions to convert periodic features into continuous features; and statistical features are constructed from historical performance data, including average efficiency, standard deviation, maximum and minimum values ​​under different conditions. The preprocessed feature vector has approximately 120 dimensions.

[0073] The parameter feature extraction model employs a deep residual network structure. The main body of the network consists of five residual blocks, each containing two fully connected layers and one skip connection. The first fully connected layer maps the input dimension from 120 to 256, while the second layer maintains the same dimension and adds the input and output through the residual connection, effectively mitigating the vanishing gradient problem in deep networks. The activation function for the residual blocks is Leaky ReLU with a slope parameter of 0.2, which, compared to traditional ReLU, preserves some gradient information in the negative region, enhancing the model's expressive power. A Dropout layer with a dropout rate of 0.3 is added after each residual block to prevent overfitting.

[0074] Following multiple residual blocks, the network introduces a feature interaction layer, which uses an attention mechanism to enable interaction between different types of features. Specifically, this is implemented as a multi-head self-attention mechanism with four attention heads, each with a dimension of 64. This design allows the model to automatically learn the interrelationships between different parameters, such as the interaction effect between the temperature coefficient and actual temperature, or the combined effect of installation angle and geographical location, significantly enhancing the model's ability to express complex parameter relationships.

[0075] The network's top layer comprises three fully connected layers with 128, 64, and 32 nodes respectively, progressively reducing the feature dimensionality. ReLU activation and batch normalization layers are used between the fully connected layers to improve training stability. The final output layer is a single neuron that outputs a predicted efficiency value in the range of 0-1 using a sigmoid activation function, representing the relative efficiency coefficient of the photovoltaic module under given parameter conditions.

[0076] The model was trained using the Adam optimizer with an initial learning rate of 0.001, employing a cosine annealing learning rate scheduling strategy, and a batch size of 128. To address the varying importance of different parameters, the loss function was designed as a weighted mean squared error, assigning higher weights to key parameters directly affecting power generation efficiency (such as temperature coefficient and shading conditions). To enhance the model's generalization ability, five-fold cross-validation was used during training, and early stopping was employed to prevent overfitting; training was stopped when the validation loss no longer decreased for 15 consecutive epochs.

[0077] To address the problem of data sparsity, especially given the limited parameter data for novel photovoltaic modules, the model incorporates transfer learning and semi-supervised learning techniques. First, the model is pre-trained on a large-scale parameter dataset to acquire general feature representation capabilities. Then, it is fine-tuned for the limited data available for specific module types. For cases where some parameters are missing, an autoencoder structure is used for parameter interpolation, effectively improving the model's robustness and adaptability. Test results show that the parameter feature extraction model can accurately predict the efficiency performance of photovoltaic modules under different conditions, with a mean absolute error of less than 3%, providing reliable module characteristic information for subsequent power generation prediction.

[0078] In this embodiment, step S3 specifically includes: training a time-series feature extraction model based on the historical data to generate time-series feature extraction results; using a spatial feature extraction model to process the spatial distribution characteristics of meteorological data to obtain spatial feature extraction results; constructing a parameter feature extraction model to process photovoltaic module parameters to generate module characteristic prediction results; and using an ensemble learning model to fuse the time-series feature extraction results, spatial feature extraction results, and module characteristic prediction results to obtain the preliminary prediction results.

[0079] like Figure 3 As shown, in the process of building a time series deep learning model, the system builds a deep learning model specifically for processing time series based on historical data, including an LSTM (Long Short-Term Memory) model to capture long-term dependencies, a TCN (Temporal Convolutional Network) model to extract multi-scale time features, and a Transformer model to process long-distance dependencies in the sequence.

[0080] For the spatial feature extraction model of meteorological data, the system constructs a convolutional neural network (CNN) model to process the spatial distribution features of meteorological data. It uses multiple convolutional layers and pooling layers to extract local and global spatial features, paying particular attention to the spatial variation patterns of meteorological elements that have a significant impact on photovoltaic power generation, such as cloud cover distribution and pressure field.

[0081] In the stage of building a deep regression model for component characteristic parameters, the system constructs a deep regression model, using the specifications, installation conditions and historical performance data of photovoltaic modules as inputs, to predict the power generation efficiency under given meteorological conditions, taking into account the impact of factors such as module type and temperature coefficient on power generation efficiency.

[0082] For model specialization training, the system trains dedicated models for different prediction time scales (short-term, medium-term, long-term) and different weather types (sunny, cloudy, rainy, etc.), and each model achieves optimal prediction performance under specific conditions.

[0083] In terms of building an ensemble learning framework, a Stacking-based ensemble learning framework is designed, using Gradient Boosting Tree (GBDT) as a meta-learner to integrate the prediction outputs of various basic models, learn the optimal model weight allocation, and form the final fusion prediction result.

[0084] The ensemble learning framework is one of the core components of this system. It is responsible for integrating the outputs of the temporal feature extraction model, spatial feature extraction model, and parametric feature extraction model to generate the final prediction result. The system adopts a stacking ensemble strategy, treating each basic model as a first-order learner. It integrates the prediction results of each model through a meta-learner (second-order learner), automatically learns the optimal model combination weights, fully leverages the advantages of each model, and improves the overall prediction accuracy.

[0085] The implementation of the ensemble learning framework is divided into two phases. In the first phase, the historical dataset is divided chronologically into a training set (70%), a validation set (20%), and a test set (10%). Each base model is trained independently using the training set, and then predictions for each model are generated on the validation set. These predictions will serve as the training input for the meta-learner. To avoid information leakage, the system employs a K-fold cross-validation strategy (K=5) to generate validation set predictions, ensuring that the data used to train the meta-learner has not been directly learned by the base models.

[0086] The meta-learner employs Gradient Boosting Decision Tree (GBDT), specifically implemented using the XGBoost algorithm. The XGBoost model is configured with 500 trees, a maximum depth of 5, a learning rate of 0.03, regularization parameters L1 of 0 and L2 of 1, and row and column sampling rates of 0.8 to enhance model robustness. The input features of the meta-learner include the predicted outputs of each base model, as well as some original features such as time features (time of day, season), key meteorological parameters (irradiance, temperature), and system state parameters (e.g., component temperature, system power), forming a feature vector of approximately 25 dimensions.

[0087] For different prediction timescales (short-term, medium-term, and long-term) and different weather types (sunny, cloudy, rainy, etc.), the system trained multiple sets of dedicated meta-learners. These dedicated meta-learners automatically adjust the weight contributions of each base model under different scenarios through adaptive feature importance analysis. For example, in short-term predictions, the time-series feature model receives a higher weight; in areas with variable weather, the spatial feature model contributes more; and in newly installed or significantly aging systems, the parametric feature model has a more prominent influence.

[0088] To improve the robustness of the ensemble model, a dynamic model selection mechanism is introduced. Based on historical prediction performance, the system constructs a model performance evaluation matrix, recording the prediction error statistics of each model under different conditions. During the prediction phase, the system queries this matrix based on current conditions (such as weather conditions, time characteristics, etc.) to dynamically allocate the optimal model combination weights for the current prediction task, and can even completely exclude poorly performing models to avoid their negative impact.

[0089] The ensemble model training employs a two-stage strategy. The first stage freezes the parameters of each base model, training only the meta-learner using mean squared error as the loss function, updating parameters with the Adam optimizer, and setting the learning rate to 0.001 for 100 epochs. The second stage uses end-to-end fine-tuning, simultaneously updating the top-level parameters of both the meta-learner and each base model, using a smaller learning rate (0.0001) for fine-tuning to further improve the overall model performance. To prevent overfitting, an early stopping strategy is implemented during training; training stops when the validation set loss no longer decreases for 10 consecutive epochs.

[0090] The ensemble learning framework also includes an uncertainty estimation module that generates confidence intervals for predictions using the Monte Carlo Dropout method (maintaining Dropout activation during the inference phase). The system performs 50 Monte Carlo samplings, calculates the mean and standard deviation of the prediction results, and generates upper and lower bounds for the 95% confidence intervals, providing a risk assessment basis for scheduling decisions.

[0091] Test results show that, compared to a single optimal model, the ensemble learning framework reduces prediction error by 21.4% and improves robustness to sudden weather changes by 32.7%. Especially when dealing with abnormal weather conditions and changes in equipment status, the ensemble model demonstrates significant advantages, automatically adjusting the weight contributions of each base model to maintain high prediction accuracy. The final prediction results output by the ensemble learning framework include predicted power generation, prediction confidence intervals, and uncertainty estimates for each time point, providing a comprehensive basis for subsequent scheduling strategy formulation.

[0092] Step S4: Based on the historical operating data of the photovoltaic system, establish a photovoltaic module aging analysis model, calculate the performance correction factor based on the photovoltaic module aging analysis model, correct the preliminary prediction results, and obtain the corrected prediction results.

[0093] In this embodiment, historical power generation data of the photovoltaic modules is first analyzed to establish a preliminary performance degradation model. The system segments the historical data according to the module's usage time and compares the power generation efficiency under similar weather conditions within each time period, plotting the module's power generation efficiency over time. By fitting these curves, the system obtains a preliminary performance degradation trend. To improve the accuracy of the analysis, the system also introduces a time-weighted mechanism, assigning higher weights to recent data, making the model more accurately reflect the current state of the modules.

[0094] Meanwhile, a theoretical aging mechanism model was established based on the physical characteristics and electrochemical reaction principles of photovoltaic modules. This model considers the main factors affecting photovoltaic module aging, including cumulative ultraviolet radiation dose (leading to yellowing of encapsulation materials and decreased light transmittance), the number and magnitude of temperature cycles (causing solder joint fatigue and microcracks), alternating damp heat (accelerating metal electrode corrosion), and mechanical loads (such as the stress effects of wind and snow loads on the module). Based on the module's material properties and structural design, a quantitative relationship model between these factors and module performance degradation was established.

[0095] This paper integrates data-driven historical performance analysis results with theoretical aging mechanism models to construct a comprehensive component aging model. The integration process employs a Bayesian framework, using the theoretical model as the prior and historical data analysis results as observations. A more accurate aging model is obtained by maximizing the posterior probability. This dual-model fusion method overcomes the shortcomings of purely theoretical models that may ignore the influence of the actual operating environment and purely data-driven models that may lack physical interpretation, resulting in a comprehensive aging model that is both theoretically sound and consistent with actual operating data.

[0096] Based on a comprehensive aging model, the module performance correction factor at the current moment is calculated. The system inputs the actual usage time of the photovoltaic module and cumulative operating environment parameters (such as cumulative radiation, historical temperature distribution, etc.) into the aging model to obtain the performance ratio of the current module relative to its initial state, which serves as the performance correction factor. This factor reflects the degree of performance degradation caused by aging, and its value typically ranges from 0.7 to 1.0, decreasing with prolonged usage time.

[0097] Finally, the performance correction factor is applied to the preliminary prediction results generated in step S3 to correct the predicted values. Specifically, the system multiplies the preliminary predicted power generation by the performance correction factor to obtain the corrected prediction value considering the aging effect. This correction mechanism significantly improves the accuracy of the prediction, especially for photovoltaic systems that have been operating for many years; the prediction deviation can be reduced by more than 30% after considering the aging effect. Simultaneously, the system also calculates the uncertainty range based on the aging model, providing a confidence interval for the prediction results and enhancing the reliability of the prediction.

[0098] In this embodiment, step S4 specifically includes: analyzing historical power generation data and operating environment data of photovoltaic modules based on the historical operating data of the photovoltaic system, and generating performance degradation analysis results; analyzing and processing environmental factors to obtain environmental impact analysis results; and fusing the performance degradation analysis results and the environmental impact analysis results to generate the aging analysis model of the photovoltaic modules.

[0099] In the historical performance data analysis process, the historical power generation data of photovoltaic modules are analyzed, and combined with factors such as environmental conditions and operating time, the trend of module performance changes over time is identified, and a preliminary performance degradation model is established.

[0100] For the establishment of the aging mechanism model, a theoretical model of module aging is established based on the physical characteristics and electrochemical reaction principle of photovoltaic modules, taking into account the influence mechanism of factors such as ultraviolet radiation, temperature cycling, and alternating damp heat on cells, encapsulation materials and connectors.

[0101] In the dual-model fusion stage, the data-driven historical performance analysis results are integrated with the theoretical aging mechanism model to establish a comprehensive module aging model. This model can predict the performance degradation curves of photovoltaic modules under different service years and different operating environments.

[0102] For the calculation of dynamic correction factor, the performance correction factor at the current moment is calculated based on the actual usage time and working environment of the photovoltaic module. This factor reflects the degree of degradation of the module relative to its initial performance and can be used to correct the prediction results.

[0103] In the prediction result correction stage, a dynamic correction factor is applied to the prediction result generated in step S3 to correct the prediction value so that it more accurately reflects the actual power generation capacity after considering the component aging effect.

[0104] The photovoltaic module aging analysis model quantitatively calculates performance degradation curves and correction factors by analyzing historical power generation data and environmental conditions, combined with the physical characteristics of the modules. In its implementation, the system first constructs a historical performance degradation dataset. For photovoltaic systems operating for more than one year, the system extracts performance data from multiple time points (e.g., one sampling point per quarter) from the historical database, with each sampling point containing at least 15 days of power generation data. To ensure data comparability, the system selects data points that closely approximate standard test conditions (irradiance between 800-1000 W / m²). 2 Within the range of component temperature (25±5℃), the actual power generation efficiency of the component under these conditions is calculated and compared with the efficiency at the time of initial installation to preliminarily quantify the performance degradation.

[0105] To accurately separate aging effects from other influencing factors, a multi-factor correction mechanism was implemented. For temperature effects, the system calculates a temperature correction coefficient based on the component's temperature coefficient (typically between -0.3% / ℃ and -0.5% / ℃) and the deviation between the actual temperature and the standard test temperature (25℃). For the nonlinear effects of irradiance, the system uses response curves based on component type to standardize the efficiency under different irradiance intensities. For the effects of spectral variations, the system estimates the efficiency changes caused by spectral shifts based on atmospheric transparency and air quality index data. Through these corrections, the system separates the influence of environmental factors from the pure aging degradation effect.

[0106] Based on corrected historical performance data, aging degradation curves for the modules are constructed. The system attempts various degradation models, including linear degradation, a two-stage degradation model (rapid initial degradation followed by slow degradation), and an exponential degradation model. The model with the smallest fitting error is selected as the degradation model for the module by fitting historical data points using the least squares method. For most crystalline silicon modules, the two-stage degradation model typically provides the best fit, exhibiting a relatively rapid degradation rate (approximately 1.5-2.5% / year) in the first 2-3 years, followed by a slower, more stable degradation (approximately 0.5-0.7% / year). The system establishes a separate degradation model for each photovoltaic module to accurately capture its unique aging characteristics.

[0107] Simultaneously, an accelerated environmental aging model was established to quantify the accelerating effect of extreme environments on component aging. The main accelerating factors considered included: high-temperature exposure (cumulative duration of temperatures exceeding 70°C), temperature cycling (cumulative number of days with daily temperature variations exceeding 40°C), humid and hot conditions (cumulative duration of temperatures exceeding 60°C and relative humidity exceeding 85%), and ultraviolet radiation (cumulative ultraviolet radiation dose). Based on environmental data monitored on-site, the system calculated the cumulative values ​​of these accelerating factors and, using an empirical model established based on accelerated aging test data, estimated the impact of environmental acceleration on component lifespan, adjusting the parameters of the basic degradation model accordingly.

[0108] When calculating the performance correction factor in real time, the system comprehensively considers the component's usage time, accumulated environmental stress, and recent performance test data. The specific steps are as follows: First, the system calculates the component's operating time (in years) based on its initial usage date; second, it obtains the theoretical degradation rate corresponding to that time point from the aging degradation model; then, it calculates the environmental acceleration factor based on accumulated environmental stress, which is typically between 0.9 and 1.2, with a value greater than 1 indicating that environmental conditions have accelerated the aging process; finally, it performs a Bayesian update based on the most recent measured performance data to obtain the corrected degradation rate. The performance correction factor is calculated as the theoretical remaining performance ratio, i.e., one minus the corrected degradation rate, and its value is typically between 0.7 and 1.0, decreasing as the component's usage time increases.

[0109] Different degradation model parameters are used for different types of photovoltaic modules. Monocrystalline silicon modules typically have an initial degradation rate of about 2% / year and a long-term stable degradation rate of about 0.5% / year; polycrystalline silicon modules have an initial degradation rate of about 2.5% / year and a long-term stable degradation rate of about 0.6% / year; thin-film modules (such as amorphous silicon, CIGS, CdTe, etc.) have a higher initial degradation rate of about 3-4% / year, but some types may exhibit extremely low long-term degradation rates after the initial degradation. The system sets initial parameters for different types of modules based on data provided by the module manufacturer and industry standard test results, and then continuously updates and adjusts these parameters based on actual operating data.

[0110] A performance correction factor is applied to the preliminary prediction results generated in step S3 to correct the predicted values. Specifically, the preliminary predicted power generation is multiplied by the performance correction factor to obtain the corrected prediction value considering the aging effect. This correction mechanism significantly improves the accuracy of the prediction, especially for photovoltaic systems that have been operating for many years; the prediction deviation can be reduced by more than 30% after considering the aging effect. Simultaneously, the system also calculates the confidence interval of the correction factor based on the uncertainty of the aging model parameters, providing a reliable estimate of the error range for the prediction results.

[0111] The aging model is updated monthly, reassessing degradation curve parameters and environmental acceleration factors based on newly added performance and environmental data to ensure the model accurately reflects the latest state of the components. Specifically, when the system detects significant anomalies in component performance (such as a sharp drop in efficiency in a short period), a special check process is triggered to distinguish whether it is due to accelerated aging, temporary performance degradation, or component failure, and the correction strategy is adjusted accordingly. This adaptive mechanism enables the aging model to accurately track dynamic changes in component performance, providing a continuously accurate basis for correction of the prediction system.

[0112] Step S5: Based on the corrected prediction results, perform short-term, medium-term and long-term power generation predictions respectively to generate multi-timescale prediction results.

[0113] In this embodiment, a short-term forecast is first performed to generate a photovoltaic power generation forecast for the next 24 hours. The short-term forecast is granular on an hourly basis, primarily focusing on the changing trend of the daily power generation curve and peak power. Real-time acquired meteorological data (including cloud cover, irradiance, temperature, etc.) and historical power generation data from the past 24 hours are input into a dedicated short-term forecast model. This model is updated at a high frequency (every 15 minutes) to ensure that the forecast results reflect changes in meteorological conditions in a timely manner. Particular attention is paid to the accuracy of forecasts during sunrise and sunset and under conditions of rapid cloud changes, providing users with accurate short-term power generation expectations and supporting the formulation of daily power generation plans. The short-term forecast results are presented in the form of a power curve, and provide confidence intervals for the hourly forecast values, reflecting the range of forecast uncertainty.

[0114] Next, a medium-term forecast is executed to generate a 7-day photovoltaic power generation forecast. The medium-term forecast is performed on a daily basis, primarily analyzing the changing patterns and periodic characteristics of daytime power generation. The system combines short-term weather forecast data provided by meteorological departments (such as 7-day weather conditions and cloud cover forecasts) with historical power generation data for the same period, inputting them into a dedicated medium-term forecast model. This model is updated daily, focusing on the impact of weather system changes on power generation. The system also improves forecast accuracy by analyzing periodic patterns in historical data (such as differences in power generation and consumption between weekdays and weekends). The medium-term forecast results are presented as a daily power generation bar chart, with the probability of sunny / rainy weather and the range of uncertainty for each day's forecast marked, providing users with a weekly planning reference.

[0115] It also performs long-term forecasts, generating photovoltaic power generation projections for the next 1-3 months. These long-term forecasts are granular on a weekly basis, primarily assessing monthly total power generation and seasonal trends. The system integrates long-term climate trend forecasts from meteorological departments, historical data for the same period, and seasonal variation patterns, inputting them into a dedicated long-term forecast model. This model is updated weekly, paying particular attention to climate change characteristics during seasonal transitions. Long-term forecasts consider more macroeconomic factors, such as seasonal variations in solar irradiance and monsoon climate characteristics, and are calibrated using multi-year historical data for the same period. The long-term forecast results are presented in the form of weekly and monthly cumulative power generation, providing comparative analysis with historical data to help users develop energy plans and economic assessments.

[0116] During the prediction process at each time scale, the dynamic correction factor generated in step S4 is used to correct the predicted values ​​for aging effects. Although the prediction models at different time scales have different focuses and update frequencies, they all share the correction results of the photovoltaic module aging model, ensuring that the predicted values ​​accurately reflect the actual performance status of the module.

[0117] Finally, consistency checks and uncertainty assessments are performed on the forecast results at each time scale. Consistency checks ensure that the predicted values ​​for short-term, medium-term, and long-term forecasts remain reasonably consistent within overlapping time periods, avoiding logical contradictions in the forecast results. Uncertainty assessments utilize methods such as Monte Carlo simulations to calculate forecast intervals at different confidence levels and generate a complete forecast report including the forecast median, upper and lower bounds, and probability distribution. The system presents these multi-time-scale forecast results in a visual manner, intuitively demonstrating the trend of power generation changes over time and the range of uncertainty, providing comprehensive information support for subsequent dispatch decisions.

[0118] In this embodiment, step S5 specifically includes: processing real-time meteorological data using a short-term forecasting model to generate short-term power generation forecasts; analyzing short-term meteorological forecast data based on a medium-term forecasting model to generate medium-term power generation forecasts; and processing seasonal variation data using a long-term forecasting model to generate long-term power generation forecasts.

[0119] During the short-term forecasting process, based on real-time meteorological data and recent historical power generation data, the short-term forecasting model constructed in step S3 is used to generate hourly power generation forecasts for the next 24 hours, with a focus on intraday trends and peak power.

[0120] For medium-term forecast execution, combining short-term weather forecasts and historical data for the same period, the medium-term forecast-specific model constructed in step S3 is used to generate daily power generation forecasts for the next 7 days, with a focus on analyzing the daily power generation variation patterns and periodic characteristics.

[0121] In the long-term forecast execution phase, long-term meteorological trends, seasonal variations, and historical annual data are integrated. Using the long-term forecast-specific model built in step S3, weekly power generation forecasts for the next 1-3 months are generated, with a focus on evaluating total monthly power generation and seasonal trends.

[0122] For the correction of prediction results, the dynamic correction factor generated in step S4 is applied to the prediction results at each time scale to correct for aging effects and improve prediction accuracy.

[0123] During the uncertainty assessment phase, the system performs confidence interval analysis on the prediction results at each time scale, calculates the uncertainty range of the predicted values, and generates a complete prediction report that includes the predicted median, upper and lower bounds, and probability distribution.

[0124] Step S6: Based on the multi-timescale prediction results and combined with electricity market data, generate grid interaction schemes and energy storage control strategies to form a photovoltaic power generation dispatch strategy.

[0125] In this embodiment, the electricity load curve is first predicted based on historical electricity consumption data. The system collects historical electricity consumption data from users and combines it with time characteristics (such as weekdays / rest days, seasonal factors), weather factors (such as temperature affecting air conditioning load), and user behavior patterns (such as work and rest schedules, electricity consumption habits) to construct a load prediction model. This model uses a combination of deep learning and time series analysis to generate electricity load prediction curves for various future time periods, accurately reflecting the changing trends on the electricity demand side. The system also provides uncertainty estimates for the predicted load curve to address the randomness of electricity consumption behavior.

[0126] Next, based on electricity pricing policies and load forecasting results, the optimal grid interaction scheme is calculated. The system takes into account local electricity pricing policies (including feed-in tariffs, peak-valley time-of-use tariffs, tiered tariffs, etc.), generation forecasting results generated in step S5, and load forecasting curves. An optimization algorithm is then used to calculate the optimal grid interaction strategy for each time period. For periods where the feed-in tariff is higher than the consumption tariff, the system tends to sell electricity to the grid; for peak tariff periods, the system prioritizes meeting its own electricity demand and reduces electricity purchases from the grid; for off-peak tariff periods with insufficient generation, the system increases the proportion of electricity purchased from the grid. By solving a mixed-integer linear programming problem, the system obtains the optimal grid interaction power for each time period, forming a detailed generation and consumption scheme.

[0127] For photovoltaic systems equipped with energy storage, the energy storage charging and discharging schedule is further optimized. The system considers technical constraints such as the capacity, charging and discharging efficiency, and cycle life characteristics of the energy storage devices. Combining power generation forecast curves, load forecast curves, and electricity price information, a dynamic programming algorithm is used to determine the optimal energy storage scheduling scheme. When there is a large peak-valley electricity price difference, the system charges during off-peak hours and discharges during peak hours to achieve peak-valley arbitrage. During periods of excess power generation, excess electricity is stored for subsequent use, improving self-consumption rate. In cases of grid curtailment, excess power is absorbed through energy storage, reducing curtailment losses. The system also considers the lifespan management of the energy storage devices, controlling the depth and frequency of charging and discharging to extend the lifespan of the energy storage system.

[0128] The system also provides intelligent management of controllable loads. For adjustable electrical equipment on the user side (such as electric vehicle chargers, water heaters, and air conditioners), the system formulates intelligent electricity consumption plans based on predicted power generation curves and electricity price information. The system adjusts controllable loads to peak photovoltaic power generation periods as much as possible, increasing the self-consumption ratio and reducing grid interaction demands. This demand-side response mechanism effectively smooths the electricity load curve, alleviates the grid's peak-shaving pressure, and saves users on electricity costs.

[0129] Finally, the system integrates grid interaction schemes, energy storage control strategies, and load management plans to form a unified dispatch strategy matrix. This matrix contains detailed operating instructions for future time periods, covering control parameters such as inverter power settings, energy storage system charging and discharging power, and controllable load switching status. The system also provides expected benefit analysis, including electricity cost savings calculations, grid-connected electricity revenue, and self-consumption rate, providing users with clear economic benefit expectations. The system displays dispatch strategies and expected effects through a visual interface, and provides explanations for strategy optimization, enhancing users' understanding and trust in the automatic dispatch system.

[0130] In this embodiment, step S6 specifically includes: using historical electricity consumption data to predict the electricity load curve and obtaining the electricity demand forecast result; calculating the grid interaction period according to the electricity price policy and the electricity demand forecast result, and generating a power generation consumption scheme; optimizing the energy storage system charging and discharging plan based on the power generation consumption scheme to form the photovoltaic power generation dispatch strategy.

[0131] Specifically, in the process of electricity load forecasting, the system predicts the electricity load curves for various time periods based on historical electricity consumption data and user behavior patterns, combined with weather factors and time characteristics, providing demand-side information for energy dispatch.

[0132] For the analysis of grid interaction strategies, the system calculates the optimal grid interaction scheme for different time periods based on electricity pricing policies, on-grid electricity prices and peak-valley price differences, combined with the generation forecast results and electricity load forecast results from step S5, including the self-consumption ratio and the timing of surplus electricity being fed into the grid.

[0133] In terms of energy storage system optimization and control, if the system is equipped with energy storage devices, the optimal charging and discharging strategy of the energy storage system is calculated based on power generation forecast, load forecast and electricity price information to balance peak-valley differences, improve self-consumption rate, and consider battery life optimization.

[0134] For controllable load management, the system formulates intelligent power consumption plans based on the predicted power generation curve for adjustable electrical equipment (such as electric vehicle chargers, water heaters, etc.), and adjusts the controllable load to the peak power generation period as much as possible to increase the self-consumption ratio.

[0135] In the integrated dispatch strategy generation stage, the system integrates the analysis results of grid interaction, energy storage control and load management to form a unified dispatch strategy matrix, which includes detailed operation instructions and expected benefit analysis for each future time period, providing clear guidance for system execution.

[0136] Step S7: Collect actual operating data of the photovoltaic system, calculate the deviation between the predicted result and the actual value, and generate prediction deviation data; based on the prediction deviation data, use an online learning method to update the prediction model parameters and improve the prediction accuracy.

[0137] Specifically, during the execution of the scheduling instructions, the system converts the scheduling strategy generated in step S6 into equipment control instructions, which are then sent to each execution unit through the system control interface, including inverter power regulation, energy storage system charging and discharging control, and controllable load switching.

[0138] For real-time monitoring and anomaly handling, the system establishes a real-time monitoring system to continuously collect the operating status and execution effect of each device, detect abnormal situations that deviate from expectations, and make real-time adjustments according to preset emergency handling rules to ensure the safe and stable operation of the system.

[0139] In the prediction deviation analysis stage, the system compares the actual power generation with the predicted value, calculates the prediction deviation and its statistical characteristics, and analyzes the causes of the deviation, including factors such as weather forecast errors, model limitations, and the impact of sudden events.

[0140] For online model updates, the system uses incremental learning and transfer learning methods based on the prediction bias analysis results to update the prediction model's parameters online, continuously improving the model's adaptability to new data patterns.

[0141] During the overall system performance evaluation and optimization phase, the system regularly evaluates the overall system performance, calculates key indicators such as prediction accuracy, self-utilization rate, and economic benefits, and optimizes system configuration, adjusts algorithm parameters, and updates scheduling strategies based on the evaluation results, forming a closed-loop feedback optimization mechanism.

[0142] Step S8: Monitor the operating status of the photovoltaic system, detect abnormal operating conditions, and generate an abnormal status report; dynamically adjust the photovoltaic power generation scheduling strategy based on the abnormal status report to ensure the safe operation of the system.

[0143] In this embodiment, the system continuously monitors the operating parameters of the photovoltaic modules, such as temperature, voltage, and current, and sets safety thresholds and abnormal pattern recognition rules to detect system anomalies in real time. When an anomaly is detected, the system generates an anomaly report containing the anomaly type, severity, and possible causes. Based on this report, the system dynamically adjusts the aforementioned scheduling strategy, including emergency measures such as reducing inverter power output, switching the energy storage system's operating mode, or activating backup power, to ensure the safe and stable operation of the entire photovoltaic system.

[0144] like Figure 4 As shown, the present invention also provides a distributed photovoltaic power generation prediction and scheduling system, comprising:

[0145] The data acquisition and preprocessing module 101 is used to acquire multi-source meteorological monitoring data and historical operation data of photovoltaic system, and to perform data cleaning and standardization processing on the multi-source meteorological monitoring data and the historical operation data of photovoltaic system to obtain standardized training data.

[0146] The data structure construction module 102 is used to construct a linked array tree data structure of a multidimensional feature space based on the standardized training data, store and index the standardized training data, and generate a data access interface.

[0147] The model building and fusion module 103 is used to obtain historical data through the data access interface, build a time-series feature extraction model, a spatial feature extraction model and a parametric feature extraction model based on the historical data, and fuse the outputs of the time-series feature extraction model, the spatial feature extraction model and the parametric feature extraction model through an ensemble learning model to obtain preliminary prediction results;

[0148] The aging correction module 104 is used to establish a photovoltaic module aging analysis model based on the historical operating data of the photovoltaic system, calculate a performance correction factor based on the photovoltaic module aging analysis model, correct the preliminary prediction result, and obtain the corrected prediction result.

[0149] The prediction execution module 105 is used to perform short-term, medium-term and long-term power generation predictions based on the corrected prediction results, and generate multi-timescale prediction results.

[0150] The scheduling strategy module 106 is used to generate a grid interaction scheme and an energy storage control strategy based on the multi-timescale prediction results and combined with electricity market data, thereby forming a photovoltaic power generation scheduling strategy.

[0151] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predictive scheduling of distributed photovoltaic power generation, characterized in that, include: Acquire multi-source meteorological monitoring data and historical operation data of photovoltaic system, and perform data cleaning and standardization on the multi-source meteorological monitoring data and the historical operation data of photovoltaic system to obtain standardized training data; Based on the standardized training data, a linked array tree data structure of multidimensional feature space is constructed to store and index the standardized training data and generate a data access interface. Historical data is obtained using the data access interface. Based on the historical data, a time-series feature extraction model, a spatial feature extraction model, and a parametric feature extraction model are constructed. The outputs of the time-series feature extraction model, the spatial feature extraction model, and the parametric feature extraction model are fused through an ensemble learning model to obtain preliminary prediction results. Based on the historical operating data of the photovoltaic system, a photovoltaic module aging analysis model is established. Based on the photovoltaic module aging analysis model, a performance correction factor is calculated to correct the preliminary prediction results and obtain the corrected prediction results. Based on the revised prediction results, short-term, medium-term and long-term power generation predictions are performed respectively to generate multi-timescale prediction results. Based on the multi-timescale prediction results and combined with electricity market data, a grid interaction scheme and energy storage control strategy are generated to form a photovoltaic power generation dispatch strategy.

2. The method according to claim 1, characterized in that, The process involves acquiring multi-source meteorological monitoring data and historical photovoltaic system operation data, then cleaning and standardizing the multi-source meteorological monitoring data and the historical photovoltaic system operation data to obtain standardized training data, including: Irradiance, temperature, humidity, cloud cover and wind speed data are obtained from multiple meteorological monitoring stations and generated into a unified time series meteorological data through synchronous data processing. Data on power generation and equipment operation parameters are collected from the photovoltaic system, outlier detection and interpolation are performed, and cleaned historical data is generated. The unified time series meteorological data and the cleaned historical data are standardized to generate the standardized training data.

3. The method according to claim 1, characterized in that, Based on the standardized training data, a linked array tree data structure for a multidimensional feature space is constructed. The standardized training data is stored and indexed, and a data access interface is generated, including: The standardized training data is partitioned into a feature space, a tree-structured storage framework is constructed, and multi-level nodes are generated. The standardized training data is stored in the corresponding tree structure nodes, and the link relationships between the nodes are established to form a linked array tree; Design a data update algorithm to dynamically maintain the linked array tree and keep the data structure balanced.

4. The method according to claim 1, characterized in that, The process involves acquiring historical data using the data access interface, constructing a temporal feature extraction model, a spatial feature extraction model, and a parametric feature extraction model based on the historical data, and fusing the outputs of the temporal feature extraction model, the spatial feature extraction model, and the parametric feature extraction model through an ensemble learning model to obtain preliminary prediction results, including: A time-series feature extraction model is trained based on the historical data to generate time-series feature extraction results. Spatial feature extraction models are used to process the spatial distribution characteristics of meteorological data to obtain spatial feature extraction results; A parameter feature extraction model is constructed to process photovoltaic module parameters and generate module characteristic prediction results. An ensemble learning model is used to fuse the temporal feature extraction results, spatial feature extraction results, and component characteristic prediction results to obtain the preliminary prediction results.

5. The method according to claim 1, characterized in that, The process involves establishing a photovoltaic module aging analysis model based on the historical operating data of the photovoltaic system, calculating a performance correction factor based on the photovoltaic module aging analysis model, and correcting the preliminary prediction results to obtain the corrected prediction results, including: Based on the historical operating data of the photovoltaic system, the historical power generation data and operating environment data of the photovoltaic modules are analyzed to generate performance degradation analysis results. Environmental factors are analyzed and processed to obtain environmental impact analysis results; The performance degradation analysis results and the environmental impact analysis results are fused together to generate the photovoltaic module aging analysis model.

6. The method according to claim 1, characterized in that, Based on the revised prediction results, short-term, medium-term, and long-term power generation predictions are performed respectively to generate multi-timescale prediction results, including: Short-term forecasting models are used to process real-time meteorological data to generate short-term power generation forecasts. Based on the analysis of short-term weather forecast data using a medium-term forecasting model, a medium-term power generation forecast value is generated. Long-term forecasting models are used to process seasonally varying data to generate long-term power generation forecasts.

7. The method according to claim 1, characterized in that, Based on the multi-timescale prediction results and combined with electricity market data, a grid interaction scheme and energy storage control strategy are generated to form a photovoltaic power generation dispatch strategy, including: Historical electricity consumption data is used to predict the electricity load curve, resulting in electricity demand forecasts. Based on the electricity pricing policy and the aforementioned electricity demand forecast, the grid interaction period is calculated, and a power generation and consumption scheme is generated. The energy storage system's charging and discharging plan is optimized based on the power generation and consumption scheme to form the photovoltaic power generation dispatch strategy.

8. The method according to claim 1, characterized in that, Also includes: Collect actual operating data of photovoltaic systems, calculate the deviation between predicted results and actual values, and generate prediction deviation data; Based on the prediction deviation data, the prediction model parameters are updated using an online learning method to improve prediction accuracy.

9. The method according to claim 1, characterized in that, Also includes: Monitor the operating status of the photovoltaic system, detect abnormal operating conditions, and generate abnormal status reports; The photovoltaic power generation scheduling strategy is dynamically adjusted based on the abnormal status report to ensure the safe operation of the system.

10. A distributed photovoltaic power generation forecasting and scheduling system, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source meteorological monitoring data and historical operation data of photovoltaic system, and to perform data cleaning and standardization processing on the multi-source meteorological monitoring data and the historical operation data of photovoltaic system to obtain standardized training data. The data structure construction module is used to construct a linked array tree data structure of a multidimensional feature space based on the standardized training data, store and index the standardized training data, and generate a data access interface. The model building and fusion module is used to obtain historical data through the data access interface, build a time-series feature extraction model, a spatial feature extraction model, and a parametric feature extraction model based on the historical data, and fuse the outputs of the time-series feature extraction model, the spatial feature extraction model, and the parametric feature extraction model through an ensemble learning model to obtain preliminary prediction results. The aging correction module is used to establish a photovoltaic module aging analysis model based on the historical operating data of the photovoltaic system, calculate a performance correction factor based on the photovoltaic module aging analysis model, correct the preliminary prediction results, and obtain the corrected prediction results. The prediction execution module is used to perform short-term, medium-term and long-term power generation predictions based on the corrected prediction results, and generate multi-timescale prediction results. The scheduling strategy module is used to generate grid interaction schemes and energy storage control strategies based on the multi-timescale prediction results and combined with electricity market data, thus forming a photovoltaic power generation scheduling strategy.

Images