Day-ahead photovoltaic power prediction method and system based on deep noise reduction auto-encoder
By combining a deep denoising autoencoder and a random forest regression model, the problems of overfitting and low computational efficiency in high-dimensional feature processing in photovoltaic power prediction are solved, achieving high-precision and robust day-ahead photovoltaic power prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-26
- Publication Date
- 2026-04-10
AI Technical Summary
Existing photovoltaic power prediction methods are prone to overfitting, low computational efficiency, and poor generalization ability when dealing with high-dimensional meteorological features.
A deep denoising autoencoder is used to reduce the dimensionality and denoise the high-dimensional features. The random forest regression model is then used to predict the day-ahead photovoltaic power. The method involves constructing highly coupled features and using a deep denoising autoencoder for feature extraction and dimensionality reduction, followed by prediction using a random forest regression model.
It improves forecast accuracy and robustness, reduces computational complexity, enhances the ability to generalize to unseen weather patterns, and meets the timeliness requirements of day-ahead forecasts.
Smart Images
Figure FT_1
Abstract
Description
Technical Field
[0001] This invention relates to the field of new energy power prediction technology, specifically to a day-ahead photovoltaic power prediction method and system based on a deep noise reduction autoencoder. Background Technology
[0002] Photovoltaic power generation output is significantly affected by meteorological factors, exhibiting strong intermittency, randomness, and nonlinearity. Day-ahead power forecasting requires predicting the power generation curve for the next 24 hours based on numerical weather forecast data before a fixed time each day. This power forecasting is crucial for power system dispatching.
[0003] Currently, there are two main methods for predicting photovoltaic power. 1. Methods based on single deep time series models (such as LSTM and GRU); although these methods can capture historical power time series patterns, their ability to process high-dimensional, multi-source meteorological features is limited. Directly inputting high-dimensional features can easily lead to too many model parameters, heavy computational burden, and overfitting during training. Furthermore, they have weak generalization ability for unseen extreme weather patterns, such as sudden cloud cover.
[0004] 2. Methods combining traditional machine learning models (such as support vector machines and random forests) with handcrafted features; these methods typically require statistical analysis and combination of the original features. However, when the dimensionality of handcrafted features is high, the model also faces the "curse of dimensionality," i.e., data sparsity in high-dimensional space. This leads to the model performing well on offline training sets, but its prediction accuracy and robustness significantly decrease when facing complex and variable weather conditions online. Multicollinearity and noise between features can also interfere with model learning. Summary of the Invention
[0005] To address the technical problems of overfitting, low computational efficiency, and poor generalization ability to unknown weather patterns in existing day-ahead photovoltaic (PV) forecasting methods when processing high-dimensional meteorological features, this invention provides a day-ahead PV power forecasting method based on a deep denoising autoencoder. This method refines the coupling relationship between solar geometry and local weather by using high-dimensional coupled feature extraction technology, and then uses a deep denoising autoencoder to reduce the dimensionality and denoise the extracted high-dimensional features. Finally, a random forest regression model is used to achieve efficient and robust day-ahead power forecasting.
[0006] The technical solution adopted in this invention is to provide a day-ahead photovoltaic power prediction method based on a deep noise reduction autoencoder, including the following steps: S1. Collect multi-source historical data related to photovoltaic power; S2. Preprocessing of multi-source historical data; S3. Construct highly coupled features by statistically analyzing, transforming, and combining multi-source historical data to generate high-dimensional feature vectors; S4. A deep denoising autoencoder is used to reduce the dimensionality of the high-dimensional feature vector and denoise it to generate a low-dimensional feature vector. S5. Using low-dimensional feature vectors as input features, a random forest regression model is used to predict day-ahead photovoltaic power. S6. The prediction results are output as curves and data files and stored in the database.
[0007] The deep denoising autoencoder in step S4 adopts a symmetrical encoder-decoder structure. The encoder consists of three fully connected layers with the number of neurons decreasing layer by layer. Each layer is followed by a ReLU activation function and a Dropout layer. The decoder is a mirror structure.
[0008] The depth denoising autoencoder in step S4 includes, S41. Denoising pre-training: Random noise is added to the high-dimensional features input to the encoder, and the decoder outputs noisy high-dimensional features. The loss function is the mean squared error, as shown in the following formula. Formula 1 In Equation 1, MSE represents the mean square error. Representing high-dimensional features, This represents a noisy high-dimensional feature, where N represents the dimension of the high-dimensional feature.
[0009] The depth denoising autoencoder in step S4 includes, S42. Feature Extraction: After pre-training, the decoder is discarded, and only the encoder is retained. The noisy high-dimensional features are input into the encoder, and the output is a denoised low-dimensional feature containing multi-source historical data information. The calculation formula is as follows. Formula 2 In Equation 2, Representing low-dimensional features, This indicates the encoder.
[0010] The multi-source historical data in step S1 includes meteorological data, astronomical data, and historical power data of photovoltaic power plants; The meteorological data includes ground temperature, humidity, air pressure, wind speed, total precipitation, total solar radiation downward shortwave radiation, total cloud cover, low cloud cover, direct solar radiation downward shortwave radiation, and diffuse solar radiation downward shortwave radiation; the astronomical data includes solar altitude angle, solar azimuth angle, and Earth-Sun distance obtained through an astronomical calculation library; the historical power data of the photovoltaic power station includes the historical measured power values of the photovoltaic power station.
[0011] The multi-source historical data preprocessing in step S2 includes missing value handling, outlier detection and correction, and feature standardization. The missing value handling includes imputation using a dual interpolation strategy of moving average and median, and the imputed value calculation formula is as follows. Formula 3 In Equation 3, The fill value is represented by w, and w represents a sliding time window of width w at that time point. This represents the average of all non-missing observations within the sliding time window. This represents the median of all non-missing observations within the sliding time window. This represents the set of indices of all non-missing observations within a sliding time window of width w at a given time point; The outlier detection and correction includes using an isolated forest model for outlier detection, and the formula for calculating the outlier score is as follows. Formula 4 In Equation 4, The score represents the outlier score, and x represents the data sample. This represents the path length in an isolated forest tree. Denotes the standardization constant for a given number of samples. This represents the average path length of a sample across all trees; a threshold TH is set when... >TH is determined to be abnormal; For a sample x that is judged to be abnormal, the median of the normal data of its corresponding feature variable in the same season and at the same time is used for replacement and correction. The feature standardization includes using RobustScaler to standardize multi-source historical data. The formula for calculating the standardized value is as follows. Formula 5 In Equation 5, Representation of features the median of This represents the interquartile distance of feature x.
[0012] The high-dimensional feature generation process in step S3 specifically involves constructing a sliding time window of length L for each photovoltaic power prediction time t, with the previous time t-1 as the endpoint. The original multi-source historical data within this sliding time window are statistically analyzed, transformed, and combined to generate a high-dimensional feature vector.
[0013] The high-dimensional feature vector includes meteorological statistical features, solar geometric features, radiation quality features, cloud dynamic features, vertex indication features, and hysteresis features; The meteorological statistical features include calculating the maximum, minimum, mean, variance, skewness, and kurtosis of the core meteorological variables within the sliding time window; The solar geometric features include calculating the maximum value, mean, variance, and cumulative duration of solar altitude angles greater than 45 degrees within the sliding time window, as well as the sine product of solar altitude angle and solar azimuth angle; the radiation quality features include calculating the total radiation, direct radiation ratio, radiation efficiency, and radiation stability index within the sliding time window; the cloud dynamic features include extracting the rate of change of total cloud cover, cloud cover trend, cloud fluctuation frequency, and cloud comprehensive impact score; the vertex indication features include whether it is a peak power generation period, the number of hours from noon, and the number of consecutive high radiation periods; the lag features include selecting key variables and extracting their lag values at times t-1, t-2, t-4, and t-8, as well as the moving average and standard deviation of the past 2 and 4 time steps.
[0014] The random forest regression model in step S5 includes, S51. During the model training phase, the random forest constructs multiple decision trees through bootstrap sampling. When splitting a node in each decision tree, a subset is randomly selected from all low-dimensional feature vectors, and the optimal split point is chosen to minimize the mean square error. The final prediction output of the model is the average of the predictions of all decision trees. S52. In the online prediction stage, for a new prediction day, firstly, the meteorological data for the next 24 hours is acquired, and high-dimensional feature vectors are obtained according to steps S1, S2 and S3; then, according to step S4, the high-dimensional feature vectors are input into the encoder of the pre-trained deep denoising autoencoder to obtain the corresponding low-dimensional feature vectors; finally, the low-dimensional feature vectors are input into the pre-trained random forest regression model to output the power prediction sequence for the next 96 15-minute time periods.
[0015] This invention provides a system for a day-ahead photovoltaic power prediction method based on a deep noise reduction autoencoder, including a data processor, a data acquisition and preprocessing module that interacts with the data processor, a high-dimensional coupled feature extraction module, a feature dimensionality reduction module, a power prediction module, and an output and storage module; The data acquisition and preprocessing module is used to acquire data from the NWP interface, the astronomical computing library and the power plant monitoring system, and to perform data cleaning, anomaly handling and standardization. The high-dimensional coupled feature extraction module has built-in feature engineering logic. Based on the configured sliding window size and feature rules, it is used to receive the preprocessed data output by the data acquisition and preprocessing module and output the transformed high-dimensional feature vector. The feature dimensionality reduction module has a built-in encoder part of a pre-trained deep denoising autoencoder, which is used to receive the high-dimensional feature vector output by the high-dimensional coupled feature extraction module and output the corresponding low-dimensional robust feature vector. The power prediction module has a built-in trained random forest regression model, which is used to receive the low-dimensional robust feature vector output by the feature dimensionality reduction module and output a sequence of photovoltaic power prediction values for the next 24 hours. The output and storage module is used to receive the prediction results output by the power prediction module, output them in the form of curves and data files, and store them in the database for system use. The data processor is used for scheduling and data transmission between modules.
[0016] The beneficial effects of this invention are that it provides a day-ahead photovoltaic power prediction method based on a deep denoising autoencoder. By introducing a deep denoising autoencoder to perform dimensionality reduction and denoising processing on high-dimensional meteorological features, dimensionality reduction can be achieved without sacrificing the accuracy of high-dimensional features. This maps high-dimensional feature representations in the data space to low-dimensional feature representations. During unsupervised learning pre-training, it enables feature reconstruction from massive high-dimensional photovoltaic power feature data to low-dimensional robust data, and reduces the impact of feature selection on power prediction. This invention has the following advantages: 1. High prediction accuracy: By constructing refined high-dimensional coupled features, the nonlinear interaction between meteorological and spatiotemporal factors is fully explored; then, a deep denoising autoencoder is used for dimensionality reduction and denoising, retaining the core information with the most predictive value and filtering out noise and redundancy, so that the subsequent random forest regression model can learn in a cleaner and more compact feature space, thereby significantly improving the prediction accuracy, especially in the prediction of weather change points.
[0017] 2. Strong robustness and generalization ability: The unsupervised pre-training of the deep denoising autoencoder enables it to learn the inherent robust representation of the data and is not sensitive to noise and slight perturbations in the input data. Combined with the anti-overfitting characteristics of the random forest regression model itself, the entire prediction framework has a stronger generalization ability to unseen weather patterns and good online prediction stability.
[0018] 3. High computational efficiency: By using a deep denoising autoencoder to reduce high-dimensional features (e.g., 158-dimensional) to low-dimensional features (e.g., 16-dimensional), the computational complexity and memory consumption of the subsequent random forest regression model are greatly reduced, shortening the model training and online prediction time and meeting the timeliness requirements of current-day prediction.
[0019] 4. Highly practical: This method has a clear process, a high degree of modularity, and is easy to implement and deploy. Attached Figure Description
[0020] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0021] like Figure 1As shown, this invention provides a day-ahead photovoltaic power prediction method based on a deep noise-reducing autoencoder, comprising the following steps: S1. Collect multi-source historical data related to photovoltaic power, including meteorological data, astronomical data, and power data; Meteorological data: provided by numerical weather prediction (NWP), including surface temperature (t2m), humidity (rh2m), air pressure (pres), wind speed (ws10), total precipitation (tp), downward shortwave total solar radiation (ssrd), total cloud cover (cldt), low cloud cover (cldl), downward shortwave direct solar radiation (bdsf_ave), and downward shortwave diffuse solar radiation (ddsf_ave). Astronomical data: The solar altitude angle (sun_altitude), solar azimuth angle (sun_azimuth), and Earth-Sun distance (sun_distance) are obtained through an astronomical computing library. Power data: including historical measured power values (P) of the photovoltaic power station.
[0022] S2. Multi-source historical data preprocessing, including missing value handling, outlier detection and correction, and feature standardization; Missing values are handled by using a double interpolation strategy based on moving average and median. For feature sequences Set the sliding time window size to w, preferably w=4, corresponding to 1 hour of data, then fill the value... The calculation formula is as follows: Formula 3 In Equation 3, The fill value is represented by w, and w represents a sliding time window of width w at that time point. This represents the average of all non-missing observations within the sliding time window. This represents the median of all non-missing observations within the sliding time window. This represents the set of indices of all non-missing observations within a sliding time window of width w at a given time point; Outlier detection and correction: An isolation forest model is used for outlier detection. For the sample, its anomaly score The calculation formula is as follows: Formula 4 In Equation 4, The score represents the outlier score, and x represents the data sample. This represents the path length in an isolated forest tree. Denotes the standardization constant for a given number of samples. This represents the average path length of a sample across all trees; a threshold TH is set when... >TH is determined to be abnormal; For a sample x that is judged to be abnormal, the median of the normal data of its corresponding feature variable in the same season and at the same time is used for replacement and correction. Feature standardization is employed to reduce the impact of outliers and data scaling on the model. RobustScaler is used to standardize the multi-source historical data. The formula for calculating the standardized value is as follows. Formula 5 In Equation 5, Representation of features the median of This represents the interquartile distance of feature x.
[0023] S3. Construct highly coupled features by statistically analyzing, transforming, and combining multi-source historical data to generate high-dimensional feature vectors; To refine the modeling of the nonlinear coupling relationship between meteorological data and solar position, a sliding time window of length L is constructed for each predicted time t, with the previous time t-1 as the endpoint. The optimal value is L=8, or 2 hours. Statistical transformations and combinations are then performed on the original multi-source data within this sliding time window to generate high-dimensional feature vectors. For example, 158 dimensions specifically include the following six categories of features. Meteorological statistical characteristics (e.g., 24-dimensional): Calculate the maximum, minimum, mean, variance, skewness, and kurtosis of the core meteorological variables (GHI-related radiation, temperature, humidity, and wind speed) within the sliding time window; Solar geometric characteristics: Calculate the maximum value, mean, variance, and cumulative duration of solar altitude angles greater than 45 degrees for the solar altitude angle sequence within the sliding time window, as well as the sine product characteristics of solar altitude angle and solar azimuth angle; Radiation quality characteristics: Calculate the total radiation, direct radiation ratio, radiation efficiency (ratio of actual radiation to theoretical maximum radiation at solar altitude angle) and radiation stability index (based on moving standard deviation) within the sliding time window. Cloud dynamic characteristics: Extract the rate of change of total cloud cover, cloud cover trend (slope of linear fitting), cloud fluctuation frequency (standard deviation), and cloud comprehensive impact score (combining changes in cloud cover, cloud height, and radiation). Peak indicator characteristics include whether it is a peak power generation period (e.g., 10:00-14:00), the number of hours from noon, and the number of consecutive high radiation periods; Lag characteristics: Select key variables (such as ssrd, t2m, rh2m, cldt, ddsf_ave, bdsf_ave, etc.) and extract their lag values at times t-1, t-2, t-4, and t-8, as well as their moving averages and standard deviations over the past 2 and 4 time steps.
[0024] The above feature extraction process is achieved through feature engineering, which couples information from different scales and sources to form a feature representation that can comprehensively characterize the short-term weather evolution, solar position, and their interaction before the prediction time.
[0025] S4. Feature dimensionality reduction and denoising based on deep denoising autoencoders The N-dimensional high-dimensional features obtained in step S3 Input the encoder part of a pre-trained deep denoising autoencoder, and output the reduced M-dimensional low-dimensional features. , For example, N=158, M=16.
[0026] The deep denoising autoencoder network structure adopts a symmetrical encoder-decoder structure. The encoder consists of three fully connected layers with the number of neurons decreasing layer by layer, for example, N → 64 → 32 → M. Each layer is followed by a ReLU activation function and a Dropout layer with a dropout rate of 0.1 to prevent overfitting. The decoder is a mirror structure, for example, M → 32 → 64 → N.
[0027] Denoising pre-training, during the training phase, involves processing the high-dimensional features of the input. Random noise, such as Gaussian noise or random masking of some feature values, is added to form a noisy input. .
[0028] The training objective is to make the decoder output... As close to the original clean as possible The loss function uses mean squared error, and the formula is as follows. Formula 1 In Equation 1, MSE represents the mean square error. Representing high-dimensional features, This represents a noisy high-dimensional feature, where N represents the dimension of the high-dimensional feature.
[0029] After feature extraction and pre-training, the decoder is discarded, retaining only the encoder. The high-dimensional features of the new data are input into the encoder, and the output vector represents the robust, denoised, and dimensionally reduced features. The calculation formula is as follows. Formula 2 In Equation 2, Representing low-dimensional features, This indicates the encoder.
[0030] S5. Using the low-dimensional features obtained in step S4 For input, the measured value of photovoltaic power at the corresponding time. To achieve this goal, a random forest regression model RFR is trained, and the trained random forest regression model RFR is used to predict day-ahead photovoltaic power.
[0031] RFR model training phase: Random Forest constructs multiple decision trees using the bootstrap method. Each decision tree, when splitting at a node, selects from all available data. A subset of features is randomly selected, and the optimal split point is chosen to minimize the mean squared error. The final prediction output of the model is the average of all decision tree predictions.
[0032] RFR Online Forecasting Phase: For a new forecast date, the photovoltaic power generation at 96:00 AM the following day is predicted. First, the meteorological data for the next 24 hours provided by the NWP is acquired, and preprocessing and feature extraction are performed according to steps S1, S2, and S3 to obtain high-dimensional features. Then, following step S4, the high-dimensional features are input into the pre-trained deep denoising autoencoder. The encoder part obtains the corresponding low-dimensional features. Finally, following step S5, the low-dimensional feature is input into the pre-trained random forest regression model, which outputs a power prediction sequence for the next 96 15-minute time intervals (i.e., 24 hours). .
[0033] S6. The prediction results are output in the form of curves and data files and stored in the database for system access.
[0034] This invention employs high-dimensional coupled feature construction, which is a multi-dimensional and refined feature engineering strategy based on sliding windows and integrating meteorological statistics, solar geometry, radiation quality, cloud dynamics, vertex indication, and hysteresis features.
[0035] This invention adopts a hybrid model architecture of DAE dimensionality reduction and denoising + RFR regression prediction. It uses unsupervised DAE to compress and denoise high-dimensional features, extracts robust low-dimensional representations, and then uses supervised RFR for accurate prediction in a two-stage process.
[0036] This invention employs a data preprocessing strategy for photovoltaic power prediction, including dual interpolation of moving average and median, and outlier handling methods based on isolation forest and replacement with historical median for the same period.
[0037] This invention provides a system for a day-ahead photovoltaic power prediction method based on a deep noise reduction autoencoder. The system includes a data processor, a data acquisition and preprocessing module that interacts with the data processor, a high-dimensional coupled feature extraction module, a feature dimensionality reduction module, a power prediction module, and an output and storage module. The data acquisition and preprocessing module is used to acquire data from the NWP interface, astronomical computing library, and power plant monitoring system, and to perform data cleaning, anomaly handling, and standardization. The high-dimensional coupled feature extraction module has built-in feature engineering logic, which, according to the configured sliding window size and feature rules, receives the preprocessed data output by the data acquisition and preprocessing module and outputs the transformed high-dimensional feature vector. The feature dimensionality reduction module has built-in encoder part of a pre-trained deep denoising autoencoder, which receives the high-dimensional feature vector output by the high-dimensional coupled feature extraction module and outputs the corresponding low-dimensional robust feature vector. The power prediction module has built-in a pre-trained random forest regression model, which receives the low-dimensional robust feature vector output by the feature dimensionality reduction module and outputs a sequence of photovoltaic power prediction values for the next 24 hours. The output and storage module receives the prediction results output by the power prediction module, outputs them in the form of curves and data files, and stores them in the database for system calls. The data processor is used for inter-module scheduling and data transmission.
[0038] To better illustrate and facilitate understanding of the present invention, the following specific example is a prediction task performed at 18:00 on a summer day in 2025 at a photovoltaic power station in Guangxi Zhuang Autonomous Region, for further explanation: The system retrieves historical data for that day and the weather data for the following day published by NWP.
[0039] For each future prediction time, such as 12:00 the next day, the data time window of 2 hours and 8 points is traced back to the previous time of 11:45 as the endpoint. Within the window, the mean, variance, and other statistical quantities of total solar radiation at ground level are calculated, along with the characteristics of solar altitude angle and the rate of cloud cover change, generating a total of 158 feature values, constituting... ; Will Inputting a pre-trained DAE encoder with an encoder structure of 158-64-32-16 yields a 16-dimensional result. .
[0040] Will Input the pre-trained RFR model and output the power prediction value at 12:00 the next day.
[0041] Repeat steps 2-5 to complete the power prediction for 96 times from 0:00 to 23:45 the next day, forming the day-ahead prediction curve.
Claims
1. A day-ahead photovoltaic power prediction method based on a deep noise-reducing autoencoder, characterized in that: Includes the following steps, S1. Collect multi-source historical data related to photovoltaic power; S2. Preprocessing of multi-source historical data; S3. Construct highly coupled features by statistically analyzing, transforming, and combining multi-source historical data to generate high-dimensional feature vectors; S4. A deep denoising autoencoder is used to reduce the dimensionality of the high-dimensional feature vector and denoise it to generate a low-dimensional feature vector. S5. Using low-dimensional feature vectors as input features, a random forest regression model is used to predict day-ahead photovoltaic power. S6. The prediction results are output as curves and data files and stored in the database.
2. The day-ahead photovoltaic power prediction method based on a deep noise-reducing autoencoder according to claim 1, characterized in that: The deep denoising autoencoder in step S4 adopts a symmetrical encoder-decoder structure. The encoder consists of three fully connected layers with the number of neurons decreasing layer by layer. Each layer is followed by a ReLU activation function and a Dropout layer. The decoder is a mirror structure.
3. The day-ahead photovoltaic power prediction method based on a deep noise reduction autoencoder according to claim 2, characterized in that: The depth denoising autoencoder in step S4 includes, S41. Denoising pre-training: Random noise is added to the high-dimensional features input to the encoder, and the decoder outputs noisy high-dimensional features. The loss function is the mean squared error, as shown in the following formula. Formula 1 In Equation 1, MSE represents the mean square error. Representing high-dimensional features, This represents a noisy high-dimensional feature, where N represents the dimension of the high-dimensional feature.
4. The day-ahead photovoltaic power prediction method based on a deep noise reduction autoencoder according to claim 3, characterized in that: The depth denoising autoencoder in step S4 includes, S42. Feature Extraction: After pre-training, the decoder is discarded, and only the encoder is retained. The noisy high-dimensional features are input into the encoder, and the output is a denoised low-dimensional feature containing multi-source historical data information. The calculation formula is as follows. Formula 2 In Equation 2, Representing low-dimensional features, This indicates the encoder.
5. The day-ahead photovoltaic power prediction method based on a deep noise reduction autoencoder according to claim 1, characterized in that: The multi-source historical data in step S1 includes meteorological data, astronomical data, and historical power data of photovoltaic power plants; The meteorological data includes ground temperature, humidity, air pressure, wind speed, total precipitation, total solar radiation downward shortwave radiation, total cloud cover, low cloud cover, direct solar radiation downward shortwave radiation, and diffuse solar radiation downward shortwave radiation; the astronomical data includes solar altitude angle, solar azimuth angle, and Earth-Sun distance obtained through an astronomical calculation library; the historical power data of the photovoltaic power station includes the historical measured power values of the photovoltaic power station.
6. The day-ahead photovoltaic power prediction method based on a deep noise-reducing autoencoder according to claim 1, characterized in that: The multi-source historical data preprocessing in step S2 includes missing value handling, outlier detection and correction, and feature standardization. The missing value handling includes imputation using a dual interpolation strategy of moving average and median, and the imputed value calculation formula is as follows. Formula 3 In Equation 3, The fill value is represented by w, and w represents a sliding time window of width w at that time point. This represents the average of all non-missing observations within the sliding time window. This represents the median of all non-missing observations within the sliding time window. This represents the set of indices of all non-missing observations within a sliding time window of width w at a given time point; The outlier detection and correction includes using an isolated forest model for outlier detection, and the formula for calculating the outlier score is as follows. Formula 4 In Equation 4, The score represents the outlier score, and x represents the data sample. This represents the path length in an isolated forest tree. Denotes the standardization constant for a given number of samples. This represents the average path length of a sample across all trees; a threshold TH is set when... >TH is determined to be abnormal; For a sample x that is judged to be abnormal, the median of the normal data of its corresponding feature variable in the same season and at the same time is used for replacement and correction. The feature standardization includes using RobustScaler to standardize multi-source historical data. The formula for calculating the standardized value is as follows. Formula 5 In Equation 5, Representation of features the median of This represents the interquartile distance of feature x.
7. The day-ahead photovoltaic power prediction method based on a deep noise reduction autoencoder according to claim 1, characterized in that: The high-dimensional feature generation process in step S3 specifically involves constructing a sliding time window of length L for each photovoltaic power prediction time t, with the previous time t-1 as the endpoint. The original multi-source historical data within this sliding time window are statistically analyzed, transformed, and combined to generate a high-dimensional feature vector.
8. The day-ahead photovoltaic power prediction method based on a deep noise reduction autoencoder according to claim 7, characterized in that: The high-dimensional feature vector includes meteorological statistical features, solar geometric features, radiation quality features, cloud dynamic features, vertex indication features, and hysteresis features; The meteorological statistical features include calculating the maximum, minimum, mean, variance, skewness, and kurtosis of the core meteorological variables within the sliding time window; The solar geometric features include calculating the maximum value, mean, variance, and cumulative duration of solar altitude angle sequences within a sliding time window, as well as the sine product features of solar altitude angle and solar azimuth angle. The radiation quality characteristics include calculating the total radiation, direct radiation ratio, radiation efficiency, and radiation stability index within the sliding time window; the cloud dynamic characteristics include extracting the rate of change of total cloud cover, cloud cover trend, cloud fluctuation frequency, and cloud comprehensive impact score; the vertex indication characteristics include whether it is a peak power generation period, the number of hours from noon, and the number of consecutive high radiation periods; the lag characteristics include selecting key variables and extracting their lag values at times t-1, t-2, t-4, and t-8, as well as the moving average and standard deviation of the past 2 and 4 time steps.
9. The day-ahead photovoltaic power prediction method based on a deep noise reduction autoencoder according to claim 1, characterized in that: The random forest regression model in step S5 includes, S51. During the model training phase, the random forest constructs multiple decision trees through bootstrap sampling. When splitting a node in each decision tree, a subset is randomly selected from all low-dimensional feature vectors, and the optimal split point is chosen to minimize the mean square error. The final prediction output of the model is the average of the predictions of all decision trees. S51. In the online prediction stage, for a new prediction day, firstly, the meteorological data for the next 24 hours is acquired, and high-dimensional feature vectors are obtained according to steps S1, S2 and S3; then, according to step S4, the high-dimensional feature vectors are input into the encoder of the pre-trained deep denoising autoencoder to obtain the corresponding low-dimensional feature vectors; finally, the low-dimensional feature vectors are input into the pre-trained random forest regression model to output the power prediction sequence for the next 96 15-minute time periods.
10. A system based on the day-ahead photovoltaic power prediction method based on a deep noise reduction autoencoder as described in any one of claims 1-9, characterized in that: It includes a data processor, as well as a data acquisition and preprocessing module that interacts with the data processor, a high-dimensional coupled feature extraction module, a feature dimensionality reduction module, a power prediction module, and an output and storage module; The data acquisition and preprocessing module is used to acquire data from the NWP interface, astronomical computing library and power plant monitoring system, and to perform data cleaning, anomaly handling and standardization; the high-dimensional coupled feature extraction module has built-in feature engineering logic, which, according to the configured sliding window size and feature rules, is used to receive the preprocessed data output by the data acquisition and preprocessing module and output the transformed high-dimensional feature vector. The feature dimensionality reduction module incorporates a pre-trained deep denoising autoencoder, used to receive the high-dimensional feature vector output by the high-dimensional coupled feature extraction module and output the corresponding low-dimensional robust feature vector. The power prediction module incorporates a pre-trained random forest regression model, used to receive the low-dimensional robust feature vector output by the feature dimensionality reduction module and output a sequence of photovoltaic power prediction values for the next 24 hours. The output and storage module receives the prediction results output by the power prediction module, outputs them in the form of curves and data files, and stores them in the database for system use. The data processor is used for scheduling and data transmission between modules.