An artificial intelligence prediction method and system for high-altitude photovoltaic grid-connected power

Abstract

This application discloses an artificial intelligence prediction method and system for grid-connected photovoltaic power generation in high-altitude areas. Belonging to the field of new energy power generation prediction technology, this method acquires the sampling time, wind speed, ambient temperature, total horizontal irradiance, and grid-connected photovoltaic power generation of high-altitude photovoltaic power plants. It performs data cleaning and constructs time-coded, lag-featured, sliding statistical, and interactive features, generating supervised samples according to a fixed historical window. Branch implicit representations are generated through local convolutional branches, bidirectional gated recurrent branches, attention-based temporal branches, and frequency domain branches. Normalized branch weights are generated using exogenous variables at the target time and fused to obtain a deep backbone prediction value. A residual correction value is then generated through a residual enhancer and, along with the tree model branch prediction values, outputs the next-time grid-connected photovoltaic power generation prediction value according to the fusion weights determined by the validation set. This method can improve the problems of prediction bias and insufficient model adaptability under sudden operating conditions in high-altitude photovoltaic scenarios.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation prediction technology, specifically an artificial intelligence prediction method and system for photovoltaic grid-connected power generation in high-altitude areas. Background Technology

[0002] Forecasting of grid-connected power generation from photovoltaic (PV) power plants is typically used for power plant generation planning, grid dispatch optimization, energy storage coordination control, and grid-connected power generation reporting. For PV power plants in high-altitude areas, total horizontal irradiance, ambient temperature, wind speed, and diurnal and seasonal variations all affect PV output. Furthermore, under conditions such as localized cloud cover, strong irradiance changes, low temperature fluctuations, or sudden wind speed changes, grid-connected PV power generation may experience rapid increases, decreases, or peak-to-valley transitions within a short period.

[0003] Current photovoltaic (PV) grid-connected power generation prediction typically relies on historical grid-connected power generation from PV power plants, meteorological station data, inverter operation data, and grid-connected metering data for modeling. However, in actual operation, the sampling periods, recording times, and data completeness of different data sources are not entirely consistent, easily leading to issues such as duplicate records, missing samples, spiking, or unavailable fields. If this data is directly fed into the model training or online inference process, it can easily cause discontinuities in the input sequence within the historical window, or inconsistencies in data caliber between the training and inference phases, thus affecting the stability of the grid-connected power generation prediction results for the next time step.

[0004] Existing forecasting methods typically employ statistical models, single recurrent neural network models, single self-attention models, or tree models to model historical grid-connected power generation and meteorological variables. While these methods can generate short-term forecasts in general scenarios, their processing links usually perform time-series fitting using a single path. In high-altitude areas where there are sudden changes in irradiance, temperature fluctuations, abrupt changes in wind speed, or rapid cloud shadow movement, it is difficult to take into account short-term abrupt changes, changes in historical context, intraday periodic variations, and multi-day periodic variations. This can easily lead to significant forecasting biases during peak periods, trough periods, or rapid ascents.

[0005] Meanwhile, existing prediction models typically use external variables such as total horizontal irradiance, ambient temperature, and wind speed as common input features during training. This lack of effective differentiation of prediction criteria under different external conditions makes it difficult for prediction results to reflect which factor—short-term disturbances, historical trends, periodic changes, or environmental factors—is primarily affecting the current situation. For dispatching operations, predicted values ​​not only need good numerical accuracy but also good interpretability and stability under abrupt changes, enabling their use in grid connection planning, dispatch optimization, and energy storage-coordinated control systems.

[0006] Furthermore, existing models are typically used directly for online inference after training. When persistent deviations occur under different sites, seasons, or high-altitude meteorological conditions, the systematic errors in the prediction results are difficult to further compress. Different types of models also vary in their adaptability to structured meteorological features and temporal variations. Without a unified output coordination mechanism, the model's adaptability to specific sites or seasons can easily become insufficient. Therefore, in scenarios where meteorological conditions vary significantly at high altitudes, grid-connected power generation fluctuates greatly, and the quality of sampled data is unstable, improving the adaptability stability of photovoltaic grid-connected power generation prediction to sudden changes in operating conditions, periodic variations, and site differences at the next moment becomes a technical problem that needs to be solved. Summary of the Invention

[0007] This application provides an artificial intelligence prediction method and system for grid-connected photovoltaic power generation in high-altitude areas. It aims to improve the problem that single prediction models in high-altitude photovoltaic scenarios cannot simultaneously characterize local disturbances, long-range dependence, and frequency domain periodic changes. Furthermore, existing models fail to dynamically adjust the contribution relationship of different prediction branches according to exogenous variables at the target time, resulting in large prediction deviations of grid-connected photovoltaic power generation under conditions of sudden changes in irradiance, temperature fluctuations, or sudden changes in wind speed.

[0008] Firstly, this application provides an artificial intelligence prediction method for grid-connected photovoltaic power generation in high-altitude areas. The method includes: acquiring data from high-altitude photovoltaic power plants, including sampling time, wind speed, ambient temperature, total horizontal irradiance, and grid-connected photovoltaic power generation; performing duplicate record deletion, invalid value removal, outlier truncation, and missing value imputation on the data to obtain cleaned data; constructing time coding, lag features, sliding statistical features, and interaction features based on the cleaned data, and generating supervised samples labeled with the grid-connected photovoltaic power generation at the next moment according to a fixed historical window; inputting the supervised samples into the photovoltaic grid-connected power generation prediction backbone, and performing prediction through local convolutional branches and bidirectional... Gated cyclic branches, attention-based temporal branches, and frequency domain branches generate branch implicit representations. The total horizontal irradiance, ambient temperature, wind speed, and time encoding at the target time are input into the physical gating module to generate normalized branch weights. These weights are then fused with the branch implicit representations to obtain a deep backbone prediction value. The deep backbone prediction value and the feature vector corresponding to the supervised sample are input into a residual enhancer to obtain a residual correction value. The deep backbone prediction value and the residual correction value are then added to obtain a corrected prediction value. The tree model branch prediction value and the corrected prediction value are fused according to the fusion weights determined by the validation set to output the photovoltaic grid-connected power prediction value for the next time step.

[0009] Using the methods described above, this application does not merely employ a single time-series model to fit historical grid-connected power generation. Instead, based on data cleaning and a fixed historical window, it simultaneously constructs time codes, lag features, sliding statistical features, and interaction features. This allows the model input to reflect the diurnal cycle, seasonal cycle, historical output inertia, meteorological fluctuations, and the combined effects of meteorological variables in high-altitude photovoltaic power plants. Furthermore, this application utilizes local convolutional branches, bidirectional gated cyclic branches, attention-based time-series branches, and frequency domain branches to extract local disturbances, contextual time-series dependencies, long-range correlations, and frequency domain periods, respectively. A physical gating module dynamically allocates branch weights based on the total horizontal irradiance, ambient temperature, wind speed, and time codes at the target time. This allows the deep backbone prediction value to adjust the contribution ratio of different branches according to changes in the external environmental conditions at the target time.

[0010] In one possible implementation, the data is cleaned by deleting duplicate records, removing invalid values, truncating outliers, and imputing missing values. This includes: parsing the data timestamps according to the sampling time; deleting duplicate records from the same sampling time; removing invalid values ​​from records with empty fields, non-numeric fields, or those unsuitable for training; truncating outliers for wind speed, ambient temperature, total horizontal irradiance, and grid-connected photovoltaic power based on the training set quantile boundaries; and imputing missing sampling records within a fixed historical window to obtain cleaned data. The cleaned data is then used to generate the data table required for subsequent feature construction. This implementation ensures consistent data processing across the training, validation, testing, and online inference phases, reducing the interference of outlier, duplicate, and missing sampling on the prediction results.

[0011] In one possible implementation, time-based coding, lag features, sliding statistical features, and interaction features are constructed based on the cleaned data. This includes: constructing hourly sine codes, hourly cosine codes, date sine codes, and date cosine codes based on the sampling time; constructing different historical lag terms for photovoltaic grid-connected power generation, total horizontal irradiance, ambient temperature, and wind speed; calculating the moving mean and moving standard deviation under different statistical windows; and constructing interaction features between total horizontal irradiance and ambient temperature, total horizontal irradiance and wind speed, and ambient temperature and wind speed. This implementation allows the supervised samples to simultaneously include periodic positions, historical lags, window fluctuations, and coupling relationships with exogenous variables, which is beneficial for improving the model's ability to represent complex meteorological conditions at high altitudes.

[0012] In one possible implementation, the supervised samples are input into the photovoltaic grid-connected power prediction backbone. Branch latent representations are generated through local convolutional branches, bidirectional gated recurrent branches, attention-based temporal branches, and frequency domain branches. This includes: projecting the multivariate sequences in the supervised samples into a unified latent space through an input embedding layer; outputting local latent representations based on dilated convolution by the local convolutional branch; outputting temporal latent representations by the bidirectional gated recurrent branch; outputting global dependency latent representations based on self-attention by the attention-based temporal branch; and outputting frequency domain latent representations after performing a fast Fourier transform on the historical photovoltaic grid-connected power sequences by the frequency domain branch. This implementation avoids the problem that single-path models can only focus on a certain type of temporal variation, allowing the photovoltaic grid-connected power prediction backbone to simultaneously retain local variations, temporal context, long-range dependencies, and periodic spectral information.

[0013] In one possible implementation, the total horizontal irradiance, ambient temperature, wind speed, and time code at the target time are input into a physical gating module to generate normalized branch weights. The branch latent representations are then fused according to these normalized branch weights to obtain a deep backbone prediction value. This includes: combining the total horizontal irradiance, ambient temperature, wind speed, and time code at the target time into a gating input vector; inputting the gating input vector into a gating network to obtain gating values ​​corresponding to local convolutional branches, bidirectional gated recurrent branches, attention temporal branches, and frequency domain branches, respectively; performing normalization on the gating values ​​to obtain branch weights with a weight sum of 1; and weighted fusion of the latent representations of each branch according to these branch weights to output a deep backbone prediction value. This implementation allows the branch weights to be generated driven by external variables at the target time, rather than using fixed average fusion or static weight fusion, thereby improving the adaptability of the prediction process to changes in irradiance, temperature, wind speed, and time cycles.

[0014] In one possible implementation, the deep backbone prediction value and the feature vector corresponding to the supervised sample are input into a residual enhancer to obtain a residual correction value. The deep backbone prediction value and the residual correction value are then added to obtain a corrected prediction value. This includes: during the training phase, reading the difference between the supervised sample label and the deep backbone prediction value as the residual label; inputting the deep backbone prediction value and the feature vector together into a gradient boosting regression model, training the residual enhancer with the residual label as the learning target; during the prediction phase, the residual enhancer outputs the residual correction value, and the deep backbone prediction value and the residual correction value are added to obtain the corrected prediction value. This implementation allows the residual enhancer to perform secondary learning on the error portion of the deep backbone that is not fully fitted, thereby compensating for systematic biases in peak segments, valley segments, and local abrupt change segments.

[0015] In one possible implementation, the predicted values ​​of the tree model branches and the corrected predicted values ​​are fused according to the fusion weights determined by the validation set to output the predicted value of the photovoltaic grid-connected power generation at the next time moment. This includes: dividing the supervised sample set into a training set, a validation set, and a test set in chronological order; reading the feature vectors from the tree model branches and outputting the tree model branch predicted values; sequentially reading the candidate weights in the candidate fusion weight sequence on the validation set, using each candidate weight as the weight of the tree model branch predicted value, subtracting the candidate weight from 1 as the weight of the corrected predicted value, and performing a weighted sum of the tree model branch predicted values ​​and the corrected predicted values ​​to obtain the fused predicted value under the corresponding candidate weights; calculating the validation set prediction error between the fused predicted value and the validation set label; determining the candidate weights whose validation set prediction errors satisfy a preset selection rule as fusion weights, and outputting the predicted value of the photovoltaic grid-connected power generation at the next time moment based on the fusion weights. Through this implementation, the fusion ratio between the deep correction output and the tree model branch output can be selected using the validation set error, so that the final predicted value takes into account both the time-frequency expression capability of the deep multi-branch model and the fitting capability of the tree model for structured features.

[0016] In one possible implementation, the preset selection rules include minimizing the mean squared error of the validation set, minimizing the mean absolute error of the validation set, or minimizing the weighted sum of the mean squared error and mean absolute error of the validation set; the candidate fusion weight sequence includes multiple candidate weights located between 0 and 1; when the validation set prediction errors corresponding to multiple candidate weights are the same, the absolute difference between each candidate weight and 0.5 is calculated, and the candidate weight with the smallest absolute difference is selected as the fusion weight. This implementation clarifies the selection criteria for the fusion weights, avoiding arbitrary settings based on human experience.

[0017] In one possible implementation, after outputting the predicted photovoltaic grid-connected power generation value for the next time moment, the method further includes: continuously reading the latest meteorological and operational data according to a preset inference cycle, updating the fixed historical window, and regenerating the input sequence and feature vector corresponding to the supervised samples; calling the photovoltaic grid-connected power generation prediction backbone, residual enhancer, and tree model branches to output the predicted photovoltaic grid-connected power generation value for the next time moment; and pushing the predicted photovoltaic grid-connected power generation value for the next time moment to the grid-connected planning system, scheduling optimization system, or energy storage collaborative control system through a descriptive state transition interface or message queue. Through this implementation, the model prediction results can be integrated into the actual power plant scheduling operations, enabling the prediction results to be used for grid-connected planning, scheduling optimization, and energy storage collaborative control.

[0018] In one possible implementation, after pushing the predicted photovoltaic grid-connected power generation value for the next moment to the grid connection planning system, scheduling optimization system, or energy storage collaborative control system, the method further includes: reading the actual photovoltaic grid-connected power generation and corresponding predicted value at the previous moment according to a preset operating cycle, and calculating the prediction error index; when the prediction error index is greater than a preset operating threshold, triggering incremental retraining; when the prediction error of the new version model obtained by incremental retraining on the validation set is greater than the prediction error of the historical model on the validation set, or when the prediction error index of the new version model in the online observation window is greater than the preset operating threshold, rolling back the current inference version to the historical model version. This implementation can form an online maintenance link for prediction output, error monitoring, incremental retraining, and version rollback, improving the model's continuous adaptability under different seasons, different sites, and different operating conditions.

[0019] Secondly, this application provides an artificial intelligence prediction system for grid-connected photovoltaic power generation in high-altitude areas, used to execute the methods described in any of the above embodiments. The system includes a data interface module, a feature engineering module, a backbone prediction module, a residual enhancement module, and a fusion decision module. Specifically, the data interface module acquires data from high-altitude photovoltaic power plants; the feature engineering module cleans the data and constructs time-coded, lag-featured, sliding statistical, and interactive features; the backbone prediction module calls the photovoltaic grid-connected power generation prediction backbone to obtain deep backbone prediction values; the residual enhancement module outputs residual correction values ​​and obtains corrected prediction values ​​based on the deep backbone prediction values ​​and feature vectors; and the fusion decision module outputs the photovoltaic grid-connected power generation prediction value for the next time step based on the tree model branch prediction values, the corrected prediction values, and the fusion weights.

[0020] Compared with related technologies, this application has the following advantages: This application removes duplicate records, eliminates invalid values, truncates outliers, and imputes missing values ​​in high-altitude photovoltaic power station data. Based on the cleaned data, it constructs time coding, lag features, sliding statistical features, and interaction features. Then, it generates supervised samples labeled with the next moment's photovoltaic grid-connected power volume according to a fixed historical window. This ensures that the model input remains consistent in terms of time, feature, and label caliber, thereby reducing the interference of duplicate records, invalid values, outliers, and missing samples on the prediction results of the next moment's photovoltaic grid-connected power volume and improving the data consistency and usability of multi-source sampling data from high-altitude photovoltaic power stations before entering the prediction model.

[0021] This application generates branched implicit representations through local convolutional branches, bidirectional gated recurrent branches, attention-based temporal branches, and frequency domain branches. It inputs the total horizontal irradiance, ambient temperature, wind speed, and time-encoded values ​​at the target time into a physical gating module to generate normalized branch weights. This allows for the separate characterization of short-term local changes within a fixed historical window, temporal dependencies, global correlations, and periodic frequency domain features. Furthermore, it enables the contribution of different branches to the deep backbone prediction value to adjust according to changes in external conditions at the target time. This reduces prediction bias caused by the single temporal fitting path's insufficient consideration of different change patterns during peak, valley, and rapid ascent segments, and improves the prediction process's adaptability to sudden changes in irradiance, temperature fluctuations, wind speed changes, and intraday periodic positions.

[0022] This application obtains residual correction values ​​by inputting the deep backbone prediction values ​​and the feature vectors corresponding to the supervised samples into a residual enhancer, and then adds the deep backbone prediction values ​​and residual correction values ​​to obtain corrected prediction values. This allows the residual patterns formed between the deep backbone prediction values ​​and supervised sample labels during the training phase to be further learned and compensated. At the same time, by fusing the tree model branch prediction values ​​and corrected prediction values ​​according to the fusion weights determined by the validation set, the final output of the next moment's photovoltaic grid-connected power generation prediction value can simultaneously utilize the deep backbone's ability to express time-series and frequency-domain features and the tree model branch's ability to fit structured features. This reduces the systematic deviations of peak and valley positions, local abrupt change positions, and specific operating conditions under the validation set error constraint, thereby improving the stability of the prediction output and the adaptability to the current site operating conditions. Attached Figure Description

[0023] Figure 1 This application provides a flowchart for predicting grid-connected photovoltaic power generation in high-altitude areas using the PHEA-Net utility model.

[0024] Figure 2 A flowchart of data acquisition and feature engineering is provided for an embodiment of this application.

[0025] Figure 3 This is a schematic diagram of a PHEA-Net multi-branch backbone and physical gating fusion structure provided in an embodiment of this application.

[0026] Figure 4 This is a flowchart of residual enhancement and adaptive fusion provided for an embodiment of this application.

[0027] Figure 5 This is a flowchart of an online maintenance process provided for an embodiment of this application.

[0028] Figure 6 This is a schematic diagram of the overall architecture and adaptive fusion application of PHEA-Net provided in an embodiment of this application. Detailed Implementation

[0029] The technical solution of this application will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are some, but not all, embodiments of this application. Unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0030] In this application, PHEA-Net is an abbreviation for Physics-aware Hyper-fusion with Error-correction and Adaptive blending, which can be described as "Physics-aware Hyper-fusion with Error-correction and Adaptive Blending Network". BiGRU is an abbreviation for Bidirectional Gated Recurrent Unit, which can be described as "Bidirectional Gated Recurrent Unit". Transformer can be described as "Time-series Modeling Network Based on Self-Attention Mechanism". FFT is an abbreviation for Fast Fourier Transform, which can be described as "Fast Fourier Transform". XGBoost is an abbreviation for eXtreme Gradient Boosting, which can be described as "Extreme Gradient Boosting Tree Model" or "Gradient Boosting Regression Model". AdamW can be described as "Adaptive Moment Estimator Optimizer with Weight Decay". MSE is an abbreviation for Mean Squared Error, and MAE is an abbreviation for Mean Absolute Error. REST can be described as "Representational State Transition Interface", and message queue can be described as "Message Middleware for Asynchronously Transmitting Prediction Results".

[0031] In this application, Time represents the sampling time of the photovoltaic power station data; WindVel represents wind speed; TAmb represents ambient temperature; GlobHor represents total horizontal irradiance; and EGrid represents the grid-connected power generation of the photovoltaic power station. High-altitude areas can be understood as areas where the elevation of the photovoltaic power station site to be predicted is not less than a preset elevation threshold, or areas recorded as high-altitude scenarios in the power station's basic ledger, dispatch area, or operation management platform. The preset elevation threshold is determined by the power station access specifications, regional operation management rules, or historical photovoltaic power station operation and maintenance ledgers, and recorded in the site configuration table. The data interface module reads the site number, site elevation value, and high-altitude scenario marker from the site configuration table. When the site elevation value is not less than the preset elevation threshold, or the high-altitude scenario marker is valid, the corresponding photovoltaic power station is used as the processing object of the method in this application. This processing object is used for subsequent data acquisition, model training, and inference output.

[0032] In some embodiments, taking two photovoltaic power stations recorded in the site configuration table as an example, if the altitude of the first photovoltaic power station is 3200 meters and the preset altitude threshold is 2500 meters, the data interface module will identify the first photovoltaic power station as a high-altitude photovoltaic power station and write its site number into the set of sites to be predicted. If the altitude of the second photovoltaic power station is lower than the preset altitude threshold, but it is marked as a high-altitude complex meteorological scenario in the operation management platform and the high-altitude scenario is marked as valid, the data interface module will also write the second photovoltaic power station into the set of sites to be predicted. For photovoltaic power stations that do not reach the preset altitude threshold and whose high-altitude scenario is marked as invalid, the data interface module will not call the high-altitude prediction processing flow of this application. In this way, subsequent data cleaning, feature construction, and model inference can be applied only to photovoltaic power stations that meet the high-altitude processing conditions, avoiding prediction bias caused by mixed training of ordinary altitude sites and high-altitude sites.

[0033] like Figure 1 As shown in the embodiments of this application, the prediction process includes a data acquisition layer, a feature engineering layer, a multi-branch backbone layer, a physical gating layer, a residual enhancement layer, an adaptive fusion layer, and a result application layer. The layers are connected sequentially according to the data output and input relationship, with the result generated by the previous layer serving as the input for the next layer.

[0034] In some embodiments, such as Figure 6 As shown, PHEA-Net can be implemented using a holistic architecture and an adaptive fusion application approach. Figure 5 The data acquisition layer is used to read meteorological data and power plant historical data. The meteorological data includes at least wind speed, ambient temperature and total horizontal irradiance. The power plant historical data includes at least sampling time and photovoltaic grid-connected power generation. The data processing and feature engineering layer is used to perform cleaning, anomaly handling, periodic encoding, hysteresis feature generation, sliding statistical feature, interactive feature construction and feature vector aggregation on the meteorological data and power plant historical data to form input data for the multi-branch backbone network to read. Figure 6 The multi-branch backbone network in PHEA-Net is the deep backbone, which includes local convolutional branches, temporal branches, global dependency branches, and frequency domain branches. The temporal branches can be implemented using bidirectional gated recurrent branches, the global dependency branches can be implemented using attention-based temporal branches, and the frequency domain branches can use Fast Fourier Transform to process historical photovoltaic grid-connected power generation sequences. The physical gated fusion module dynamically allocates the weights of each branch based on the external variables at the target time and fuses the outputs of each branch to obtain... Figure 6 The Edeep described in the document corresponds to the deep backbone prediction value in this application. The residual enhancer reads the deep backbone prediction value and the feature matrix to calculate... Figure 6The E_residual mentioned in the document corresponds to the residual correction value in this application. The adaptive fusion output section fuses the corrected prediction value after residual enhancement, the tree model branch prediction value, and the optimized weight w to obtain the output prediction result. The results and applications section uses the output prediction result for at least one of grid connection plan application, scheduling optimization, and energy storage coordination.

[0035] It should be noted that, Figure 6 This is a simplified embodiment of the overall architecture of this application. Figure 6 In this context, "periodic encoding" corresponds to temporal encoding in this application; "feature vector convergence" corresponds to feature vector generation during the construction of supervised samples; "temporal branch" corresponds to bidirectional gated loop branch; "global dependency branch" corresponds to attention temporal branch; "Edeep" corresponds to deep backbone prediction value; "Eresidual" corresponds to residual correction value; "weight w optimization" corresponds to the process of determining fusion weights on the validation set; and "output prediction result" corresponds to the predicted photovoltaic grid-connected power generation value at the next time step. The above simplified descriptions are for illustrative purposes only. Figure 6 The structural relationships of the embodiments shown do not change the processing objects, processing order, and output results defined in each step of this application.

[0036] like Figure 2 As shown, the data interface module receives meteorological data, operational data, and grid-connected metering data respectively, and aligns the sampling time according to the Time field; the feature engineering module sequentially performs data cleaning, time encoding construction, lag feature construction, sliding statistical feature construction, and interactive feature construction, and finally outputs supervised samples.

[0037] In some embodiments, the data acquisition layer reads data from a photovoltaic power plant data acquisition system, a weather station, an inverter data interface, a grid-connected metering interface, or a dispatch data interface. Each data entry includes at least Time, WindVel, TAmb, GlobHor, and EGrid. Specifically, WindVel, TAmb, and GlobHor are collected by the power plant weather station or an external weather interface; EGrid is collected by the grid-connected metering device, the inverter monitoring system, or the power plant monitoring and data acquisition system; and Time is provided by the data acquisition system timestamp or the dispatch center timestamp. The data interface module sorts and aligns the data from each source according to the sampling time to form a raw data table. Each row in the raw data table corresponds to a sampling time, and each column corresponds to a data field.

[0038] In some embodiments, when multiple sampled values ​​exist from different data sources within the same sampling period, the data interface module generates a record according to a preset aggregation rule. The preset aggregation rule may include: taking the cumulative value of photovoltaic grid-connected power generation within the same sampling period, and taking the average, hourly value, or end-of-sampling value for wind speed, ambient temperature, and total horizontal irradiance. The preset aggregation rule is determined by the power plant data access frequency, the scheduling system prediction granularity, and the grid-connected metering method, and is recorded in the data access configuration table. During online inference, the data interface module reads the data access configuration table and generates a data record for the current sampling time according to the same rule.

[0039] In some embodiments, the feature engineering layer first performs data cleaning on the original data table. Specifically, the data interface module reads the Time field and marks records whose Time field cannot be parsed into timestamps as invalid records. For duplicate records with the same Time, one record is retained according to a preset retention rule. The preset retention rule can be to retain the record with the highest field completeness or to retain the record whose data access time is later in the sorting. After deduplication, a deduplicated data table is generated.

[0040] In some embodiments, invalid value removal includes: reading the WindVel, TAmb, GlobHor, and EGrid fields from the deduplication data table; marking the corresponding record as invalid when the field value is empty, non-numeric, or cannot be converted into a model input value; deleting invalid records from the deduplication data table to generate a valid data table. The valid data table is used for subsequent outlier truncation and missing value imputation.

[0041] In some embodiments, outlier truncation employs quantile truncation. Specifically, based on the effective data table from the training phase, the lower and higher quantile boundaries of the WindVel, TAmb, GlobHor, and EGrid fields are calculated respectively. The same quantile boundary configuration table is invoked during the training, validation, testing, and inference phases. When a field value is less than its corresponding lower quantile boundary, the field value is replaced with the lower quantile boundary; when a field value is greater than its corresponding higher quantile boundary, the field value is replaced with the higher quantile boundary. Through the above processing, a truncated data table is obtained. The quantile boundary configuration table includes field names, lower quantile boundaries, higher quantile boundaries, and generation time, used to ensure consistency in data processing across the training and inference phases.

[0042] In some embodiments, missing value imputation is used to ensure data continuity within a fixed historical window. Specifically, the truncated data table is checked for continuity based on the Time field. When there is a missing sampling time between two adjacent records, the feature engineering module inserts a blank record corresponding to the missing sampling time and imputs it based on the valid values ​​of the same field before and after the missing sampling time. If a valid value exists on only one side of the missing sampling time, the most recent valid value on that side is used for filling; if valid values ​​exist on both sides of the missing sampling time, adjacent valid value interpolation is used for filling. A cleaned data table is generated after imputation. The cleaned data table includes continuous sampling timestamps and the corresponding values ​​for each field.

[0043] In some embodiments, taking the original data table of a certain sampling day as an example, if there are two records for the sampling time of 10:15 AM on June 10, 2025, where the first record is missing the GlobHor field and the second record includes the WindVel, TAmb, GlobHor, and EGrid fields, the data interface module will retain the second record and delete the first record according to the field completeness priority rule. If the GlobHor field value in the record of 10:30 AM on June 10, 2025 is greater than the high quantile boundary formed during the training phase, the feature engineering module will replace the GlobHor field value with the corresponding high quantile boundary. If the record of 10:45 AM on June 10, 2025 is missing, but there are valid GlobHor, TAmb, WindVel, and EGrid field values ​​at both the previous and subsequent sampling times, the feature engineering module will interpolate and fill in the missing data based on the valid values ​​before and after the sampling time to generate a complete record corresponding to 10:45 AM. The processed data records are written into a cleaned data table and used in the training, validation, testing and online inference phases according to the same cleaning criteria, thereby reducing the impact of inconsistent data access criteria on the prediction results.

[0044] In some embodiments, the feature engineering layer constructs a time-encoded data table. For any sampling time, its hour value and the date sequence number within a year are read, and the hour sine code, hour cosine code, date sine code, and date cosine code are calculated respectively, thereby obtaining the time-encoded feature table. The time-encoded feature table is used to express the diurnal cycle and seasonal cycle, and is input into the subsequent model as part of the supervised samples. The hour value is denoted as... The date serial number is recorded as The number of days in a year is recorded as The time code can be obtained as follows: in, Indicates the sampling time The corresponding hourly sine code, Indicates the sampling time The corresponding hourly cosine code, together with the other, is used to characterize the position of the day-night cycle within a day; Indicates the sampling time The corresponding date sine code, Indicates the sampling time The corresponding date cosine encoding, together with the date cosine encoding, is used to characterize the position of the seasonal cycle within a year. Indicates the sampling time The corresponding hour value can range from 0 to 23; Indicates the sampling time The date sequence within a year; This indicates the number of days in the corresponding year, which can be 365 or 366. Represents pi; ( ) represents the sine function. ( () represents the cosine function. Subscript Indicates sine code, subscript Represents cosine encoding, subscript This indicates the corresponding sampling time. Through the above sine and cosine encoding, the hour value and date sequence number can be converted into continuous periodic features for subsequent construction of supervised samples and reading by the physical gating module.

[0045] In some embodiments, the feature engineering layer constructs hysteresis features. For each field in EGrid, GlobHor, TAmb, and WindVel, the field value corresponding to different hysteresis lengths before the current sampling time is read and used as the hysteresis term for the corresponding field. The hysteresis length can include 1 hour, 2 hours, 3 hours, 6 hours, 12 hours, 24 hours, 48 ​​hours, and 72 hours. If the hysteresis length before the current sampling time is insufficient, that time is not used as the starting point for supervised samples. The hysteresis feature table includes field names, hysteresis lengths, and hysteresis values ​​for use in the construction of subsequent supervised samples.

[0046] In some embodiments, the feature engineering layer constructs moving statistical features. For each field in E_Grid, GlobHor, TAmb, and WindVel, the moving mean and moving standard deviation of the field values ​​within each window are calculated, using 3 hours, 6 hours, 12 hours, 24 hours, and 48 hours prior to the current sampling time as statistical windows. The moving mean reflects short-term or medium-term level changes, and the moving standard deviation reflects the fluctuation range within the window. The generated moving statistical feature table is used together with time encoding, lag features, and interaction features to form a feature vector.

[0047] In some embodiments, the feature engineering layer constructs interactive features. For the same sampling time, the products of GlobHor and TAmb, GlobHor and WindVel, and T_Amb and WindVel are calculated respectively, and the results are used as interactive terms. The interactive terms are used to express the combined effects of irradiance and temperature, irradiance and wind speed, and temperature and wind speed on grid-connected power generation changes.

[0048] In some embodiments, the feature engineering layer constructs supervised samples according to a fixed historical window. Taking a fixed historical window of 72 hours as an example, for a target time t, the cleaned data, time encoding, lag features, sliding statistical features, and interaction features from the 72 hours prior to t are read and arranged into an input sequence in chronological order. The E_Grid corresponding to the target time t is read as the label. The sampling times in the cleaned data table that meet the historical window length requirement are traversed to generate a supervised sample set. Each supervised sample includes an input sequence, a feature vector, exogenous variables at the target time, and a label. Among them, the exogenous variables at the target time include the GlobHor, TAmb, WindVel, and time encoding at the target time.

[0049] In some embodiments, taking a fixed historical window of 72 hours and a sampling interval of 15 minutes as an example, when the target time is 12:00 PM on June 10, 2026, the feature engineering module reads continuous sampling records from the 72 hours prior to the target time and arranges them in order of time from earliest to latest to form a multivariate input sequence. The multivariate input sequence includes WindVel, TAmb, GlobHor, EGrid, time encoding, lag features, sliding statistical features, and interaction features corresponding to each sampling time. The feature engineering module simultaneously reads GlobHor, TAmb, WindVel, as well as hourly sine, hourly cosine, date sine, and date cosine codes corresponding to 12:00 PM on June 10, 2026, as exogenous variables for the target time, and reads the actual EGrid corresponding to the target time as the supervision sample label. If there is less than 72 hours of continuous data before a target time, or if a complete input sequence cannot be formed after interpolation, then that target time is not used as the starting point for supervision samples. Therefore, it can be ensured that there is a clear temporal correspondence between the input sequence, the external variables at the target time, and the labels.

[0050] In some embodiments, the fixed historical window can be 48 hours, 72 hours, or 96 hours. During the training phase, the system constructs supervised samples using candidate historical windows and calculates the prediction error on the validation set. The candidate historical window with the smallest prediction error on the validation set is selected as the fixed historical window for the current site. The determined fixed historical window is used for sample construction in the subsequent online inference phase.

[0051] like Figure 3As shown, the input embedding layer maps the multivariate sequences in the supervised samples to a unified latent space, and then inputs local convolutional branches, bidirectional gated recurrent branches, attention temporal branches, and frequency domain branches respectively. The physical gating module reads the external variables at the target time and outputs the normalized branch weights, and then performs weighted fusion of the latent representations of each branch according to the normalized branch weights.

[0052] In some embodiments, the multi-branch backbone layer includes an input embedding layer, a local convolutional branch, a bidirectional gated recurrent branch, an attention temporal branch, and a frequency domain branch. The input embedding layer reads the input sequence and projects the multivariate sequence onto a unified latent space through a linear mapping to obtain an embedding sequence. The embedding sequence serves as the input to the local convolutional branch, the bidirectional gated recurrent branch, and the attention temporal branch.

[0053] In some embodiments, the local convolutional branch reads the embedded sequence and performs convolution processing on the local changes at adjacent sampling times and interval sampling times through at least one layer of dilated convolution. The input of the dilated convolution is the time-step features in the embedded sequence, and the output is a local latent representation. The local latent representation is used to characterize high-frequency local fluctuations and rapid climbing features, and participates in weighted fusion in the physical gate layer.

[0054] In some embodiments, a bidirectional gated loop branch reads the embedded sequence, updates the hidden state according to forward and reverse time respectively, and concatenates or maps the forward and reverse hidden states to obtain a temporal hidden representation. The temporal hidden representation is used to characterize the contextual dependencies within the historical window and participates in weighted fusion in the physical gating layer.

[0055] In some embodiments, the attention-based temporal branch reads the embedded sequence and calculates the correlation between different sampling times within a historical window using self-attention. The self-attention calculation outputs a global dependency latent representation. This global dependency latent representation is used to characterize long-range temporal correlations and participates in weighted fusion in the physical gating layer.

[0056] In some embodiments, the frequency domain branch reads the historical EGrid sequence within a fixed historical window, performs a Fast Fourier Transform on the historical EGrid sequence to obtain the spectrum result, reads the amplitude vector corresponding to the main frequency from the spectrum result, and projects the main frequency amplitude vector onto the unified latent space through a mapping layer to obtain the frequency domain latent representation. The frequency domain latent representation is used to characterize the periodic spectral characteristics of grid-connected power and participates in weighted fusion in the physical gating layer.

[0057] In some embodiments, local implicit representations, temporal implicit representations, globally dependent implicit representations, and frequency domain implicit representations are mapped to the same dimension before entering the physical gating layer. If the output dimension of any branch differs from the dimension of the unified latent space, dimension alignment is performed through the corresponding linear mapping layer. After dimension alignment, the set of branch implicit representations is obtained.

[0058] In some embodiments, the physical gating layer reads the external variables at the target time and inputs them into the gating network. The gating network outputs gating values ​​corresponding one-to-one with multiple branches. The physical gating layer performs Softmax normalization on the multiple gating values ​​to obtain the branch weights, making the sum of the weights of all branches equal to 1. Taking four branches as an example, the first... The unnormalized gating values ​​corresponding to each branch are denoted as follows: , No. The weight of each branch is denoted as The branch weights are then generated as follows: Subsequently, the physical gating layer multiplies the weight of each branch with the corresponding implicit representation, and sums the results to obtain the fused implicit representation. The implicit representation of each branch is denoted as The fusion of implicit representations is denoted as The fused implicit representation is obtained as follows: The fused implicit representation outputs the deep backbone prediction value through the prediction layer. The branch weights can also serve as prediction explanation information, recording the contribution ratio of each branch to the deep backbone prediction value at the target time.

[0059] In some embodiments, taking a rapid uphill period in the morning as an example, the GloBor at the target time is significantly higher than the previous sampling time, and WindVel shows a sudden change. After reading the exogenous variables at the target time, the physical gating module can generate relatively high local convolutional branch weights, so that the local latent representation output by the local convolutional branch occupies a high proportion in the fused latent representation, thereby enhancing the responsiveness to rapid changes in adjacent sampling times. If the target time is in a stable power generation uphill phase under continuous sunny conditions, and the EGrid change curves of the same period in the most recent historical days have similar periods, the physical gating module can increase the weight of the frequency domain branch or the attention temporal branch, so that the frequency domain latent representation or the global dependency latent representation contributes more to the deep backbone prediction value. If the target time is in the recovery phase after cloud cover, the local convolutional branch, the bidirectional gated cyclic branch, and the attention temporal branch can all obtain high weights, while utilizing short-term abrupt change information and the contextual dependencies within the historical window. Through the above branch weight adjustment methods, the exogenous variables at the target time no longer participate in the prediction only as ordinary input features, but are used to dynamically change the contribution relationship of multiple branch latent representations to the deep backbone prediction value.

[0060] In some embodiments, the physical gating module can also output branch weight records. The branch weight records include the target time, the external variables at the target time, the weights of each branch, and the corresponding predicted values. The branch weight records can be synchronously written to the runtime database along with the final predicted values, for subsequent analysis of whether a given prediction time is primarily driven by local fluctuations, temporal dependencies, global dependencies, or frequency domain periods.

[0061] In some embodiments, when training the photovoltaic grid-connected power prediction backbone, the supervision sample set is divided into a training set, a validation set, and a test set in chronological order, with the proportions of the training set, validation set, and test set being 70%, 15%, and 15%, respectively. The training set is used to update model parameters, the validation set is used to select model weights, fusion weights, and early stopping times, and the test set is used to evaluate the final prediction performance of the model. Dividing the set in chronological order can prevent data leakage from future moments to historical training samples.

[0062] In some embodiments, the photovoltaic grid-connected power prediction backbone is trained using the AdamW optimizer, with the loss function being a weighted combination of mean squared error and mean absolute error. Gradient pruning, learning rate scheduling, and early stopping observation are employed during training. Gradient pruning limits the gradient magnitude of a single parameter update; learning rate scheduling adjusts the learning rate based on the training epochs or validation set error; early stopping observation stops training when the validation set error fails to improve for several consecutive epochs, and saves the model weights whose validation set error satisfies a preset selection rule. The supervised sample labels are denoted as... The predicted value of the deep backbone is denoted as The number of training samples is denoted as The mean square error weight is denoted as The loss function can then be calculated as follows: in, The value is between 0 and 1, determined by the results of the validation set error comparison, and recorded in the model training configuration table.

[0063] like Figure 4 and Figure 5 As shown, the residual enhancer reads the deep backbone prediction value and feature vector, outputs the residual correction value, and obtains the corrected prediction value; the fusion decision module reads the tree model branch prediction value and the corrected prediction value, and generates the final prediction value according to the fusion weight determined by the validation set. After the final prediction value is pushed to the grid connection planning system, the error monitoring module reads the actual grid-connected power consumption according to the operating cycle and calculates the error index. When the error index is greater than the preset operating threshold, incremental retraining or version rollback is triggered.

[0064] In some embodiments, the residual enhancement layer is executed after the training of the photovoltaic grid-connected power prediction backbone is completed. The residual enhancement layer reads the deep backbone prediction value and label corresponding to the training or validation set samples, and calculates the difference between the label and the deep backbone prediction value as the residual label. Subsequently, the deep backbone prediction value and feature vector are used as input to the residual enhancer, and the residual label is used as the residual enhancer label to train the gradient boosting regression model. After training, the residual enhancer reads the deep backbone prediction value and feature vector of the current sample and outputs the residual correction value. The residual enhancement layer adds the deep backbone prediction value and the residual correction value to obtain the corrected prediction value.

[0065] In some embodiments, taking a specific training sample as an example, the actual EGrid corresponding to this training sample is 900 kWh, and the deep backbone prediction value output by the photovoltaic grid-connected power prediction backbone is 870 kWh. The residual enhancement layer then uses 30 kWh as the residual label for this training sample. During training, the residual enhancer simultaneously reads the deep backbone prediction value, GlobHor hysteresis term, TAmb sliding statistics, WindVel sliding statistics, and interaction features of this sample to learn the deviation pattern of the deep backbone prediction value relative to the actual EGrid under similar high irradiance, low temperature, or sudden wind speed changes. In the prediction phase, if the deep backbone prediction value output by the photovoltaic grid-connected power prediction backbone for a target time is 870 kWh, the residual enhancer outputs a residual correction value of 25 kWh based on the feature vector corresponding to that target time. The residual enhancement module then adds 870 kWh to 25 kWh to obtain a corrected prediction value of 895 kWh. Therefore, the residual enhancer does not replace the deep backbone prediction again, but rather performs secondary compensation for the unfitted residual part of the deep backbone prediction value.

[0066] The residual label is denoted as The residual correction value output by the residual amplifier is denoted as... The corrected predicted value is denoted as , Indicates the sampling time The corresponding residual correction value, i.e., the error compensation amount output by the residual enhancer based on the predicted depth backbone value and the feature vector; superscript The residual correction process can be expressed as follows: (This represents the corrected prediction result.) In some embodiments, the adaptive fusion layer also trains tree model branches. The tree model branches read feature vectors and directly learn labels, outputting tree model branch predictions. The fusion decision module searches for fusion weights on the validation set. Specifically, a candidate fusion weight sequence is pre-set, including multiple candidate values ​​between 0 and 1. For each candidate fusion weight, the fusion decision module uses that candidate fusion weight as the weight of the tree model branch prediction, subtracts the candidate fusion weight from 1 as the weight of the corrected prediction, and performs a weighted sum of the tree model branch prediction and the corrected prediction to obtain the fusion prediction value under that candidate fusion weight. The tree model branch prediction value is denoted as... The candidate fusion weight is denoted as The fused predicted value is denoted as The fused prediction value is obtained as follows: The fusion decision module calculates the prediction error between the fused predicted values ​​on the validation set and the labels, and uses the candidate fusion weights whose prediction errors satisfy preset selection rules as the final fusion weights. The preset selection rules can be: minimum mean squared error of the validation set, minimum mean absolute error of the validation set, or minimum weighted sum of mean squared error and mean absolute error of the validation set. The determined fusion weights are used in the online inference phase.

[0067] In some embodiments, the online inference process is executed in a rolling manner according to a preset inference cycle. The data interface module reads the latest meteorological and operational data according to the preset inference cycle and updates and cleans the data table; the feature engineering module updates the time code, lag features, sliding statistical features, and interactive features based on the latest fixed historical window; the backbone prediction module inputs the updated input sequence into the photovoltaic grid-connected power prediction backbone to obtain the deep backbone prediction value; the residual enhancement module outputs the residual correction value based on the deep backbone prediction value and feature vector, and obtains the corrected prediction value; the fusion decision module calls the tree model branch to read the feature vector, generates the tree model branch prediction value, and reads the tree model branch prediction value, the corrected prediction value, and the fusion weight, and outputs the photovoltaic grid-connected power prediction value for the next time moment. The photovoltaic grid-connected power prediction value for the next time moment is sent to the grid connection planning system, the scheduling optimization system, or the energy storage collaborative control system as the grid connection plan recommendation value for the next time moment.

[0068] In some embodiments, the preset inference period is determined based on the photovoltaic power plant data sampling period, the granularity of grid-connected power reporting, and the dispatch system receiving period, and is recorded in the operation configuration table. For example, when meteorological data and grid-connected metering data are accessed at a 15-minute granularity, and the dispatch system receives forecast values ​​at a 15-minute granularity, the preset inference period is set to 15 minutes. The preset operation period is determined based on the actual grid-connected power feedback period and the error monitoring frequency, and is recorded in the same operation configuration table. During online operation, the system reads the preset inference period and preset operation period from the operation configuration table, and triggers forecast output and error monitoring processing respectively.

[0069] For example, during the morning hours at a high-altitude photovoltaic power station, the system reads the historical data from the past 72 hours as input. If the total horizontal irradiance rises rapidly before and after the target time, the local convolutional branch can capture the rapid rise between adjacent sampling times; if the period is still in the intraday power generation increase phase, the time code and historical EGrid sequence can reflect the diurnal cycle; if similar power generation waveforms exist at the same time in recent days, the frequency domain branch can provide a periodic reference. The physical gating module adjusts the weights of each branch based on the total horizontal irradiance, ambient temperature, wind speed, and time code at the target time, so that the final deep backbone prediction value considers both abrupt changes and intraday and multi-day cycles. The residual enhancer then compensates for the deep backbone prediction value based on historical error patterns, and the fusion decision module merges the tree model branch prediction values ​​and the corrected prediction values ​​to output the photovoltaic grid-connected power generation prediction value for the next time moment.

[0070] In some embodiments, taking a specific validation set time as an example, the predicted value of the tree model branch is 860 kWh, the corrected predicted value is 840 kWh, and when the candidate fusion weight is 0.3, the fused predicted value is 0.3 × 860 + 0.7 × 840 = 846 kWh. The system sequentially calculates the prediction error on the validation set for different candidate fusion weights. If the validation set mean square error corresponding to the candidate fusion weight of 0.3 is the smallest, then 0.3 is determined as the fusion weight. This fusion weight is written to the model version file and read by the fusion decision module during the online inference phase.

[0071] In some embodiments, the prediction system is deployed on a power plant edge server or a dispatch center. During deployment, the data interface module receives data via a REST interface or message queue; the feature engineering module performs data cleaning and feature construction in a central processing unit (CPU) environment; the backbone prediction module performs inference in a graphics processing unit (GPU) or CPU environment; the residual enhancement module and the fusion decision module output corrected prediction values ​​and final prediction values; and the results application layer writes the final prediction values ​​and branch weight information into the dispatch system or the power plant operation database.

[0072] In some embodiments, the preset retention rules are determined based on data source priority and field completeness; when multiple records correspond to the same Time, the record with the highest field completeness is retained first, and if the field completeness is the same, the record with the later data access time is retained. Quantile boundaries are determined based on the historical distribution of the corresponding fields in the training set, and the field name, low quantile boundary, and high quantile boundary are recorded in the model version file. The preset selection rules are determined based on the validation set error index, which can be the minimum mean square error of the validation set, the minimum mean absolute error of the validation set, or the minimum weighted value of both. The preset operating threshold is determined based on the power plant dispatching application error requirements, the historical online model error distribution, or the allowable error range confirmed by maintenance personnel, and is recorded in the operating configuration table. During online operation, the system reads the corresponding threshold in the operating configuration table, compares the current period error index with the threshold, and generates a retraining trigger flag when the error index is greater than the preset operating threshold. The number of times the error has occurred is used as the basis for determining the retraining trigger. Taking the operational observation window as an example, the prediction error index can be expressed as mean squared error or mean absolute error, and calculated as follows: in, Mean squared error is used to measure the average level of the squared deviation between the predicted and the actual values. Mean absolute error is used to measure the average level of absolute deviation between the predicted value and the actual value. Indicates the first The actual grid-connected photovoltaic power generation corresponding to each sample or predicted time; Indicates the first The predicted value of grid-connected photovoltaic power generation corresponding to each sample or prediction time; Indicates the first The squared prediction error for each sample; Indicates the first The absolute value of the prediction error for each sample.

[0073] In some embodiments, the operation and maintenance process includes error monitoring, drift detection, periodic retraining, and online version switching. The error monitoring module reads the actual grid-connected photovoltaic power generation and corresponding predicted values ​​at each occurrence according to a preset operating cycle, and calculates the mean square error, mean absolute error, or error index configured by the business. When the error index exceeds a preset operating threshold, the current site is marked as requiring incremental retraining. During incremental retraining, the system reads new data and updates the training sample set, retrains the photovoltaic grid-connected power generation prediction backbone, residual enhancer, and tree model branches, and redetermines the fusion weights on the validation set. During online version switching, the system retains historical model versions and records preset stability conditions in the operating configuration table. The preset stability conditions include validation set error conditions and online observation error conditions; the validation set error condition is that the prediction error of the new version model on the validation set is not greater than the prediction error of the historical model on the same validation set, or the difference between the two is not greater than the preset version error tolerance; the online observation error condition is that the prediction error index of the new version model in each operating cycle within the online observation window is not greater than the preset operating threshold. The system reads the prediction error of the new version model validation set, the prediction error of the historical model validation set, the actual grid-connected photovoltaic power and the corresponding predicted value in the online observation window, and compares the above conditions one by one. If both the validation set error condition and the online observation error condition are met, the new version model is recorded as the current inference version. If any condition is not met, the current inference version is rolled back to the historical model version, and a version rollback record including the site number, model version number, unmet conditions and rollback time is generated.

[0074] In some embodiments, taking the online operation of a photovoltaic power station as an example, the system continuously outputs the predicted photovoltaic grid-connected power generation value for the next moment according to a preset inference cycle of 15 minutes, and reads the actual EGrid for the already occurred moment according to the preset operation cycle. When the mean absolute error within the operation observation window is greater than the preset operation threshold, the error monitoring module generates an incremental retraining trigger flag. The system reads the newly added operation data and regenerates the supervised sample set, performs incremental retraining on the photovoltaic grid-connected power generation prediction backbone, residual enhancer, and tree model branches, and redetermines the fusion weights on the validation set. If the mean absolute error of the new version model on the validation set is less than that of the historical model, but the prediction error within the online observation window of the new version model is still greater than the preset operation threshold, the system will not continue to use the new version model as the current inference version, but will switch the current inference version back to the historical model version and record the rollback reason. If the new version model satisfies both the validation set error condition and the online observation error condition, the system will write the new version model into the current inference version configuration and call the new version model to output the predicted value in subsequent preset inference cycles. This approach avoids directly replacing the online model simply because the validation set error has decreased, thereby reducing the risk of unstable model updates in high-altitude, rapidly changing weather scenarios.

[0075] In some embodiments, the prediction system includes a data interface module, a feature engineering module, a backbone prediction module, a residual enhancement module, and a fusion decision module. The data interface module collects and sorts data; the feature engineering module generates cleaned data tables, time-coded data, lag features, moving statistical features, interactive features, and supervised samples; the backbone prediction module calls the photovoltaic grid-connected power prediction backbone to output deep backbone prediction values; the residual enhancement module calls the gradient boosting regression model to output residual correction values ​​and calculates corrected prediction values; and the fusion decision module calls the tree model branches to output tree model branch prediction values ​​and outputs the final prediction value based on the fusion weights.

[0076] This application also provides a computer device. The computer device includes a processor, a memory, and a communication interface. The memory stores a computer program, and the processor executes the computer program to implement the artificial intelligence prediction method for grid-connected photovoltaic power generation in high-altitude areas as described in any of the foregoing embodiments. The communication interface is used to communicate with a photovoltaic power plant data acquisition system, a meteorological data system, a dispatching system, or a message queue.

[0077] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the artificial intelligence prediction method for grid-connected photovoltaic power generation in high-altitude areas as described in any of the foregoing embodiments.

[0078] It should be noted that, in this document, the terms "comprising," "including," and any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Specific examples have been used in this document to illustrate the principles and implementation methods of the present invention. These examples are merely for the purpose of helping to understand the method and core ideas of the present invention. The above descriptions are only preferred embodiments of the present invention. It should be pointed out that, due to the limitations of written expression and the objective existence of infinite specific structures, those skilled in the art can make several improvements, modifications, or variations without departing from the principles of the present invention, and can also combine the above technical features in an appropriate manner. These improvements, modifications, variations, or combinations, or the direct application of the concept and technical solution of the present invention to other situations without modification, should all be considered within the scope of protection of the present invention.

Claims

1. An artificial intelligence prediction method for grid-connected photovoltaic power generation in high-altitude areas, characterized in that, include: Data from high-altitude photovoltaic power stations is acquired, including sampling time, wind speed, ambient temperature, total horizontal irradiance, and grid-connected photovoltaic power generation. The data is then cleaned by deleting duplicate records, removing invalid values, truncating outliers, and imputing missing values. Based on the cleaned data, time coding, lag features, sliding statistical features, and interaction features are constructed, and supervised samples labeled with the next moment's photovoltaic grid-connected power are generated according to a fixed historical window; The supervised samples are input into the photovoltaic grid-connected power prediction backbone, and branch implicit representations are generated through local convolutional branches, bidirectional gated cyclic branches, attention temporal branches, and frequency domain branches. The total horizontal irradiance, ambient temperature, wind speed, and time code at the target time are input into the physical gating module to generate normalized branch weights. The branch implicit representation is then fused according to the normalized branch weights to obtain the deep backbone prediction value. The deep backbone prediction value and the feature vector corresponding to the supervised sample are input into the residual enhancer to obtain the residual correction value, and the deep backbone prediction value and the residual correction value are added to obtain the corrected prediction value. The predicted values ​​of the tree model branches are fused with the corrected predicted values ​​according to the fusion weights determined by the validation set, and the predicted value of the photovoltaic grid-connected power generation at the next moment is output.

2. The method according to claim 1, characterized in that, The cleaned data obtained includes: The data timestamp is parsed according to the sampling time. Duplicate records with the same sampling time are deleted, and invalid values ​​are removed from records with empty fields, non-numeric fields, or records that cannot participate in training. Outlier truncation is performed on wind speed, ambient temperature, total horizontal irradiance, and grid-connected photovoltaic power based on the training set quantile boundaries; Imputation is performed on the missing sampling records within a fixed historical window to obtain cleaned data.

3. The method according to claim 1, characterized in that, The supervised samples are input into the photovoltaic grid-connected power prediction backbone, and branch latent representations are generated through local convolutional branches, bidirectional gated recurrent branches, attention-based temporal branches, and frequency domain branches, including: The multivariate sequence in the supervised sample is projected to a unified latent space by the input embedding layer; the local latent representation is output by the local convolution branch based on dilated convolution, the temporal latent representation is output by the bidirectional gated recurrent branch, the global dependent latent representation is output by the attention temporal branch based on self-attention, and the frequency domain latent representation is output by the frequency domain branch after performing a fast Fourier transform on the historical photovoltaic grid-connected power generation sequence.

4. The method according to claim 1, characterized in that, The total horizontal irradiance, ambient temperature, wind speed, and time code at the target time are input into the physical gating module to generate normalized branch weights. The branch implicit representation is then fused according to these normalized branch weights to obtain the deep backbone prediction value, including: The total horizontal irradiance, ambient temperature, wind speed, and time code at the target time are combined into a gating input vector; The gated input vector is input into the gated network to obtain the gated values ​​corresponding to the local convolutional branch, the bidirectional gated recurrent branch, the attention temporal branch, and the frequency domain branch, respectively. The gate value is normalized to obtain the branch weights with a weight sum of 1. The hidden representations of each branch are then weighted and fused according to the branch weights to output the deep backbone prediction value.

5. The method according to claim 1, characterized in that, The deep backbone prediction value and the feature vector corresponding to the supervised sample are input into the residual enhancer to obtain the residual correction value. The deep backbone prediction value and the residual correction value are then added to obtain the corrected prediction value, including: During the training phase, the difference between the supervised sample labels and the predicted values ​​of the deep backbone is used as the residual labels. The deep backbone prediction value and the feature vector are input into the gradient boosting regression model, and the residual booster is trained with the residual label as the learning target. During the prediction phase, the residual enhancer outputs a residual correction value, and the depth backbone prediction value is added to the residual correction value to obtain the corrected prediction value.

6. The method according to claim 1, characterized in that, Output the predicted photovoltaic grid-connected power generation value for the next time step, including: The supervised sample set is divided into training set, validation set and test set in chronological order; The feature vector is read from the tree model branch and the predicted value of the tree model branch is output. On the validation set, candidate weights in the candidate fusion weight sequence are read sequentially. Each candidate weight is used as the weight of the tree model branch prediction value. The value obtained by subtracting the candidate weight from 1 is used as the weight of the corrected prediction value. The tree model branch prediction value and the corrected prediction value are weighted and summed to obtain the fusion prediction value under the corresponding candidate weight. Calculate the validation set prediction error between the fused prediction value and the validation set label; determine the candidate weights whose validation set prediction errors satisfy the preset selection rules as fused weights, and output the photovoltaic grid-connected power prediction value at the next moment based on the fused weights.

7. The method according to claim 6, characterized in that, The preset selection rules include minimizing the mean squared error of the validation set, minimizing the mean absolute error of the validation set, or minimizing the weighted value of the mean squared error and mean absolute error of the validation set; the candidate fusion weight sequence includes multiple candidate weights located between 0 and 1; when the validation set prediction errors corresponding to multiple candidate weights are the same, the absolute difference between each candidate weight and 0.5 is calculated, and the candidate weight with the smallest absolute difference is selected as the fusion weight.

8. The method according to claim 1, characterized in that, After outputting the predicted photovoltaic grid-connected power generation value for the next time moment, the method further includes: The system reads the latest meteorological and operational data in a rolling manner according to a preset inference cycle, updates the fixed historical window, and regenerates the input sequence and feature vector corresponding to the supervised sample. It calls the photovoltaic grid-connected power prediction backbone, residual enhancer, and tree model branch to output the photovoltaic grid-connected power prediction value at the next moment. The photovoltaic grid-connected power prediction value at the next moment is pushed to the grid connection planning system, scheduling optimization system, or energy storage collaborative control system through the descriptive state transition interface or message queue.

9. The method according to claim 8, characterized in that, After pushing the predicted photovoltaic grid-connected power generation value for the next moment to the grid-connected planning system, dispatch optimization system, or energy storage collaborative control system, the method further includes: The actual grid-connected photovoltaic power and corresponding predicted values ​​at the time of occurrence are read according to the preset operating cycle, and the prediction error index is calculated. When the prediction error index is greater than the preset operating threshold, incremental retraining is triggered. When the prediction error of the new version model obtained by incremental retraining is greater than the prediction error of the historical model on the validation set, or when the prediction error index of the new version model in the online observation window is greater than the preset operating threshold, the current inference version is rolled back to the historical model version.

10. An artificial intelligence prediction system for grid-connected photovoltaic power generation in high-altitude areas, characterized in that, The system for performing the method according to any one of claims 1 to 9, the system comprising: The data interface module is used to acquire data from high-altitude photovoltaic power stations; the feature engineering module is used to clean the data and construct time coding, hysteresis features, sliding statistical features, and interaction features. The backbone prediction module is used to call the photovoltaic grid-connected power prediction backbone to obtain the deep backbone prediction value; the residual enhancement module is used to output the residual correction value and obtain the corrected prediction value based on the deep backbone prediction value and the feature vector. The fusion decision module is used to output the predicted value of the photovoltaic grid-connected power generation at the next moment based on the predicted value of the tree model branch, the corrected predicted value, and the fusion weight.

Images