A Photovoltaic Power Generation Prediction Method Based on Pattern-Aware Adaptive Deep Learning

Abstract

This invention discloses a photovoltaic power generation prediction method based on pattern-aware adaptive deep learning, comprising: S1, extracting daily features based on historical operating data of photovoltaic power plants, automatically identifying and classifying at least two operating modes, and mapping the identified operating mode labels to all data time points within the natural day corresponding to the daily features; S2, for the operating modes identified in step S1, performing signal decomposition and feature extraction on the corresponding photovoltaic power time series, calculating signal complexity features based on the decomposition results, and constructing a feature set adapted to each mode by combining mode-specific external meteorological and time features; S3, constructing a unified deep learning prediction model containing a pattern embedding layer based on a bidirectional long short-term memory network, loading specific hyperparameters for the identified operating modes, and inputting the constructed feature set adapted to the operating modes and the mode labels into the prediction model for training, and outputting power prediction results.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation, and more specifically to a photovoltaic power generation prediction method based on pattern-aware adaptive deep learning. Background Technology

[0002] Amid the global push to address climate change and accelerate energy structure transformation, photovoltaic (PV) power generation, as a core component of renewable energy, is experiencing unprecedented exponential growth. However, the large-scale, high-proportion integration of PV power also poses serious challenges to the safety, stability, and economic operation of the power system. The core challenge stems from the inherent intermittency and volatility of PV output, as well as its extreme sensitivity to complex and variable weather conditions. Therefore, developing advanced predictive models capable of accurately capturing and adapting to the complex dynamic characteristics of PV is of crucial scientific significance and practical value for improving renewable energy integration rates and ensuring the reliability and economic efficiency of the power system.

[0003] Existing photovoltaic power prediction methods can be broadly categorized into three types: physical models, statistical models, and data-driven machine learning / deep learning models. The accuracy of physical models highly depends on detailed meteorological data and precise system parameters, and is often limited in practical applications due to parameter uncertainties. Traditional statistical models, such as ARIMA (Autoregressive Integral Moving Average), are often inadequate in characterizing the inherently strong nonlinear and non-stationary characteristics of photovoltaic power generation sequences.

[0004] With technological advancements, deep learning techniques, such as Long Short-Term Memory (LSTM) networks, Transformer architectures, and their hybrid variants, have become mainstream in photovoltaic (PV) forecasting. Simultaneously, signal decomposition techniques, exemplified by Improved Ensemble Empirical Mode Decomposition (ICEEMDAN), are widely used in data preprocessing. However, these advanced techniques generally share a common problem in their application: they mostly operate in a "mode-agnostic" manner, employing fixed models and parameters to process all data throughout the year, lacking the ability to adapt to the actual operating modes of PV power plants. This "mode-agnostic" paradigm leads to the following key technological bottlenecks:

[0005] Coarseness of model segmentation: Existing studies generally treat seasonality by directly ignoring seasonal differences or relying solely on fixed calendar dates for simple segmentation. This approach struggles to accurately capture the dynamic model boundaries determined by both actual weather conditions and operational characteristics, resulting in insufficient model generalization ability during model transition periods or under special weather conditions.

[0006] Non-adaptive nature of feature engineering: Most studies tend to use fixed signal decomposition parameters and feature extraction strategies for all data, ignoring the significant differences in frequency composition, noise level and signal complexity of photovoltaic signals under different operating modes. This limits the depth and accuracy of extracting effective information from the raw data.

[0007] Lack of pattern awareness at the model level: Standard deep learning models function as "general sequence processors". Their internal structure generally lacks the ability to perceive the current pattern and cannot dynamically adjust their feature learning and prediction strategies according to the identified operating pattern. This results in limited robustness when facing high fluctuations or atypical working conditions.

[0008] Therefore, there is an urgent need for a new technical solution that can overcome the above-mentioned bottlenecks and solve the problems of insufficient refined seasonal adaptation and model pattern perception capabilities in photovoltaic power generation forecasting. Summary of the Invention

[0009] To address the technical problems mentioned in the background section regarding existing photovoltaic (PV) prediction models' lack of pattern awareness and inability to adapt to diverse PV power generation operation modes, this invention provides a PV power prediction method based on pattern-aware adaptive deep learning. Through data-driven pattern recognition, pattern-specific feature engineering, and a unified architecture-based end-to-end parameter specialization strategy, the prediction accuracy and robustness are significantly improved.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a photovoltaic power generation prediction method based on pattern-aware adaptive deep learning, comprising the following steps:

[0011] S1. Based on historical operating data of photovoltaic power plants, daily characteristics are extracted, and at least two operating modes are automatically identified and divided. The identified operating mode labels are mapped to all data time points within the natural day corresponding to the daily characteristics. Based on the daily operating characteristics, a Gaussian Mixture Model (GMM) is used to perform cluster analysis on the annual operating days of the photovoltaic power plant. To scientifically determine the optimal number of operating modes K, this study evaluates the results by calculating the Bayesian Information Criterion (BIC) scores under different numbers of clusters. The analysis results show that when K=4, the BIC score reaches the optimum, and the clustering results have clear physical meaning and can effectively distinguish the four key operating modes. Therefore, the optimal number of clusters K=4 is finally determined.

[0012] S2. For each operating mode identified in step S1, independently perform signal decomposition and feature extraction on the corresponding photovoltaic power time series, calculate the signal complexity features based on the decomposition results, and construct a feature set that is compatible with each mode by combining mode-specific external meteorological and time features; combine mode-specific external meteorological and time features to generate mode-specific input features that match the characteristics of each mode.

[0013] S3. Construct a unified deep learning prediction model architecture with a pattern embedding layer based on BiLSTM-Transformer; for each operating mode identified in step S1, independently load its optimized exclusive hyperparameters, and input the feature set and mode label of the mode into the model instantiated based on the unified architecture for training, and finally output the predicted value of photovoltaic power generation.

[0014] It should be noted that in step S1 of this invention, daily features are used for clustering and a unique macroscopic pattern label is assigned to the data for the whole day. This is a unique technical design that decouples 'macroscopic operational background perception' from 'microscopic time series dynamic prediction'.

[0015] Compared to traditional methods that attempt to mix and process signals of different time scales in a single model, the advantages of this invention are:

[0016] Stability of macroscopic identification: Step S1 first uses GMM clustering to identify stable and interpretable daily operating conditions (i.e., macroscopic background) on a daily scale, avoiding the contamination of macroscopic condition judgment by intraday instantaneous disturbances. 2. Clear division of responsibilities: The deep learning model in step S3 is responsible for specifically processing and capturing specific intraday dynamic changes (such as power fluctuations caused by instantaneous cloud cover) under the premise of known macroscopic pattern background. This decoupled design of division of responsibilities is the key technical prerequisite for the high-precision and high-robustness prediction of this invention.

[0017] In the above-mentioned pattern-aware adaptive deep learning photovoltaic power generation prediction method, the specific implementation of step S1 includes:

[0018] S11. Daily Feature Extraction: The historical operation data of the photovoltaic power station is aggregated into daily data, and multiple daily features covering meteorological conditions and irradiance patterns are extracted. From the high-frequency (15-minute interval) raw photovoltaic operation data, daily features are extracted for each natural day. The daily features cover two dimensions: meteorological conditions and irradiance patterns, including total daily irradiance, total daily direct irradiance, average daily temperature, average cloud cover, irradiance centroid time, and fluctuation indicators, etc.

[0019] S12. Feature Standardization: To eliminate the impact of differences in the dimensions and numerical ranges of different daily features on subsequent cluster analysis, the Z-score standardization method is used. Specifically, for any feature element in the daily feature vector of any given day, subtract the mean of that feature across all date samples, and then divide by the standard deviation of that feature across all date samples to obtain the standardized feature element. (where the feature dimension D=11), its standardized vector The first in element The calculation is as follows:

[0020] ;

[0021] in, For the standardized first Each feature element; For the original first Each feature element; For the first The mean of each feature across all date samples; For the first The standard deviation of each feature across all date samples;

[0022] S13. Clustering and Pattern Determination: The standardized daily features are clustered using a Gaussian mixture model clustering algorithm. By calculating the Bayesian Information Criterion (BIC) under different K values, the K value with the lowest or best BIC score is selected as the optimal number of patterns. Dates with similar features are grouped into the same cluster, and each cluster is defined as a running pattern. The identified running pattern labels are then mapped to all data time points within the natural day corresponding to the daily features.

[0023] In the above-described pattern-aware adaptive deep learning photovoltaic power generation prediction method, the specific implementation of step S2 includes:

[0024] S21. Mode-specific signal decomposition: For photovoltaic power time series under each operating mode... The improved complete set empirical mode decomposition (ICEEMDAN) method is used independently to decompose it into a series of intrinsic mode functions (IMFs) and a residual term. Its expression is:

[0025] ;

[0026] in, It is a pattern The number of IMFs obtained from the time series decomposition of photovoltaic power. The decomposition yields the first... One Intrinsic Mode Function (IMF) For the residual terms; each intrinsic mode function (IMF) represents the local oscillation characteristics of the signal at different frequency scales;

[0027] S22, Adaptive Alignment: To unify the input dimensions of subsequent models, the decomposed... One IMF component and one residual term (total) The components are forcibly recombined into a fixed number of N component sequences (N=9 in this embodiment); the recombination rules are: a) retain the first N-1 (i.e., 8) IMF components. b) Retain one residual term and place it in the Nth component position; c) If the actual IMF quantity If there are more than N-1 IMF components, the extra (i.e., the Nth and subsequent) IMF components are summed point by point and added to the Nth component (residual term) to ensure dimensionality uniformity. The advantage of this 'accumulation' strategy (rather than 'discarding' strategy) is that it guarantees the integrity of the signal decomposition information. That is, the sum of the N component sequences (N-1 IMFs and 1 composite residual term) can still accurately reconstruct the original photovoltaic power time series mathematically, avoiding information loss caused by discarding high-frequency components. More importantly, as described in subsequent step S23, the feature extraction of this invention is only performed on the first N-1 IMF components, and the Nth 'composite residual component' does not participate in subsequent feature calculations. Therefore, this alignment operation ensures information integrity without introducing confusion into the feature set of the input model.

[0028] S23. Signal complexity feature extraction for pattern specialization: For each pattern, independently calculate the fuzzy entropy value of the N-1 IMF component sequences (excluding residual components) after decomposition and alignment. And these calculated FE values ​​themselves (e.g.) , ,..., These features are directly used to characterize the complexity of the signal and incorporated into the final pattern specialization input feature set.

[0029] S24. External Feature Selection and Processing for Model Specialization: a) Define and filter relevant weather feature subsets (such as GHI, DNI, Temperature, etc.) for each model; b) Extract cyclic coding features based on timestamps, including sine and cosine transforms of hours, days of the year, and days of the week; c) Create lag sequences of past multiple time steps for the target power and selected weather features.

[0030] In the above-mentioned pattern-aware adaptive deep learning photovoltaic power prediction method, the deep learning prediction model used in step three is a pattern-aware BiLSTM-Transformer hybrid model, and the specific implementation of its pattern-aware mechanism is as follows:

[0031] 1) The feature sequence received by the model contains a special pattern ID feature (integer 0-3), which clearly indicates the operating mode to which the current data point belongs;

[0032] 2) The pattern ID is input into an embedding layer (torch.nn.Embedding). This layer internally maintains a trainable embedding matrix, whose function is to map the input pattern ID (an integer) representing the category into a low-dimensional, dense vector of real numbers. This process can be represented by the following formula:

[0033] ;

[0034] ;

[0035] in: In time step The pattern ID (an integer, such as 0, 1, 2, or 3). The lookup function for the embedding layer (i.e.) It retrieves the corresponding row vector from the embedding matrix based on the input integer ID. It is the output pattern embedding vector, which is a... Dense vectors in space. It is the preset embedding dimension, a hyperparameter that determines the richness of the pattern representation;

[0036] 3) During training, the model automatically learns and optimizes these pattern embedding vectors by backpropagating the loss of the prediction task;

[0037] 4) Learned pattern embedding vectors Compared with other feature vectors at this time step The features are concatenated to form an enhanced feature vector. As input to subsequent BiLSTM layers:

[0038] ;

[0039] in: In time step The enhanced final input vector. Represents the time step The vector consisting of all the original input features has a dimension of . ;

[0040] 5) Parameter specialization mechanism under unified architecture:

[0041] Step S3 of this invention employs a deep adaptation strategy of 'unified architecture and parameter specialization'.

[0042] First, the pattern embedding layer serves as a contextual prior: the model incorporates a pattern embedding layer at the input. During independent training, this layer maps fixed pattern labels into low-dimensional, dense embedding vectors. These vectors act as a contextual prior, providing the network with baseline feature distribution information based on current physical conditions (such as 'summer peak season'), thus accelerating model convergence.

[0043] Secondly, deep adaptation of the parameter space: Combining hyperparameter optimization with pattern specialization, this invention does not only divide the data at the data level, but also independently optimizes each pattern in the entire parameter space (including embedded vector values, network weights, and hyperparameter configurations). This design retains the uniformity and standardization of the model architecture (facilitating engineering deployment), while ensuring the model's ability to perfectly fit the dynamic characteristics of specific patterns through independent training.

[0044] The aforementioned pattern-aware adaptive deep learning photovoltaic power prediction method also includes a pattern-specific hyperparameter optimization step. Specifically, it uses the Optuna framework based on Bayesian optimization to independently search for hyperparameters for each identified operating mode. The optimized hyperparameter space covers ICEEMDAN decomposition parameters, fuzzy entropy calculation parameters, and BiLSTM-Transformer model structure, training, and regularization parameters.

[0045] This step is performed before the final model training, and its purpose is to independently determine an optimal set of hyperparameter combinations for each identified operating mode. This step specifically includes:

[0046] 1. Independent Optimization: Using Bayesian optimization frameworks such as Optuna, hyperparameter search is performed in an independent "Study" for each operating mode (e.g., 'summer peak', 'inefficient / cloudy'). 2. Full-Link Parameter Space: The optimized hyperparameter space includes not only the deep learning model structure and training parameters (e.g., number of BiLSTM units, number of attention heads, learning rate), but also the feature engineering parameters described in this invention (e.g., ICEEMDAN decomposition parameters and fuzzy entropy calculation parameters). 3. Overfitting Prevention Mechanism: To prevent hyperparameters from overfitting to the training set during optimization, an independent validation set is used for evaluation. For example, the data for each mode is divided into a training set (e.g., the first 80%) and a validation set (e.g., the last 20%) according to time sequence. The goal of the optimization framework is to minimize the model's prediction error (e.g., RMSE) on this independent validation set, thereby finding the parameter combination with the best generalization ability.

[0047] The output of the optimization step is a configuration dictionary containing the optimal hyperparameters specific to each mode. This configuration set is called in the subsequent final model training and evaluation steps.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] In terms of pattern recognition, this invention employs a data-driven Gaussian Mixture Model (GMM) clustering method. Based on the daily characteristics of meteorological conditions and irradiance patterns, it identifies four key operating modes with different physical meanings and power generation characteristics (spring and autumn medium efficiency, autumn and winter medium efficiency, summer peak, and low efficiency / cloudy days) from the annual operating data. This method overcomes the limitations of traditional seasonal division based on a fixed calendar and achieves the capture of dynamic operating characteristics of photovoltaic power plants.

[0050] This invention employs daily granularity for macroscopic pattern segmentation in pattern recognition. By leveraging the stability of daily aggregated features, it ensures the robustness and physical interpretability of pattern recognition, avoiding interference from intraday instantaneous disturbances on macroscopic pattern recognition (such as 'summer peak day'). The complex intraday dynamic capture task is uniformly handled by the subsequent BiLSTM-Transformer model. This architecture, which first defines the macroscopic operating conditions and then resolves the microscopic dynamics, solves the technical problem of inappropriate pattern perception granularity in existing technologies.

[0051] For the different identified operating modes, this invention combines signal decomposition (ICEEMDAN) and feature extraction (FE) strategies optimized independently for each mode to achieve adaptive feature engineering that specializes in the mode. The framework demonstrates superior prediction performance in all four modes, and maintains high accuracy even in the most challenging high-fluctuation modes (such as "summer peak") and inefficient modes (such as "inefficient / cloudy"), significantly outperforming non-adaptive benchmark methods. This strongly demonstrates the high accuracy, robustness and excellent generalization ability of the proposed framework.

[0052] Regarding model design and performance, this invention proposes a deep adaptive strategy of 'full-link parameter specialization'. Unlike traditional one-size-fits-all single models that struggle to accommodate multimodal data distributions, this invention uses mode-specific hyperparameter optimization (HPO) to independently determine the optimal parameter combination, including signal decomposition, feature extraction, and model structure, for each operating mode, and independently trains dedicated model instances. This design ensures that the model can achieve an extreme fit to the dynamic characteristics of different modes, significantly improving the system's robustness under complex meteorological conditions while maintaining prediction accuracy.

[0053] In terms of theoretical and applied value, the pattern-aware adaptive deep learning framework proposed in this paper integrates data-driven pattern discovery, pattern-specialized signal processing and feature engineering, and a prediction model with a pattern embedding mechanism, forming a complete end-to-end solution. Research results show that this framework provides an effective new paradigm for solving the long-standing seasonal adaptation problem in photovoltaic power generation forecasting. It has significant theoretical value and engineering application prospects for improving the grid's ability to accommodate and dispatch high proportions of renewable energy, and promoting the safe, stable, and economical operation of the power system. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the overall prediction process of the present invention.

[0055] Figure 2 This is a schematic diagram of the correlation heatmap analysis of daily candidate features before clustering in this invention;

[0056] Figure 3 This is a schematic diagram of the BIC and AIC score analysis results used to select the optimal number of clusters K when using GMM clustering in this invention;

[0057] Figure 4 This is a schematic diagram showing the monthly distribution of the number of days for the four new operating modes identified in this invention.

[0058] Figure 5 This is a comparison chart of the average daily power curves of the four new operating modes identified by this invention;

[0059] Figure 6 A schematic diagram of the internal structure of the pattern-aware BiLSTM-Transformer hybrid prediction model designed for this invention;

[0060] Figure 7 This is a comparison chart of the prediction results of each mode during the final evaluation of this invention;

[0061] Figure 8 This is a scatter plot of the predicted and actual values ​​for the entire year, representing the combined values ​​of all models, used for the final evaluation of this invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0063] like Figure 1As shown, a pattern-aware adaptive deep learning method for predicting photovoltaic power generation mainly includes the following steps:

[0064] S1. Based on historical operating data of photovoltaic power plants, Gaussian mixture model (GMM) cluster analysis is used to extract features on a daily basis and automatically identify and classify four photovoltaic operating modes with different physical meanings and operating characteristics (i.e., spring and autumn medium efficiency, autumn and winter medium efficiency, summer peak, and low efficiency / cloudy days) (their monthly distribution is as follows). Figure 4 As shown, its average daily power curve is as follows: Figure 5 As shown in the figure, a unique pattern label is assigned to the data at each time point. The specific implementation method is as follows:

[0065] S11. Daily Feature Extraction: First, the raw high-frequency (15-minute interval) photovoltaic operation data is cleaned and interpolated. Then, daily features are extracted from the preprocessed data for each natural day. The daily features cover two dimensions: meteorological conditions and irradiance patterns, including total daily irradiance, total daily direct irradiance, average daily temperature, average cloud cover, irradiance centroid time, and fluctuation indicators.

[0066] Before clustering, to eliminate the impact of collinearity among features on the clustering effect, it is necessary to select the most representative feature subset. This invention performs correlation analysis on all the aforementioned non-power diurnal features (including meteorological conditions and irradiance morphology features).

[0067] Figure 2 A visual heatmap for correlation analysis, based on Figure 2 Based on the analysis results, this invention identifies and eliminates highly redundant features (i.e. Figure 2 Feature pairs with an absolute correlation value close to 1.0. For example, among complete candidate features, features such as daily average horizontal irradiance (GHI_mean), daily maximum total horizontal irradiance (GHI_max), daily average direct normal irradiance (DNI_mean), and daily maximum direct normal irradiance (DNI_max) show strong correlations with daily total horizontal irradiance (GHI_sum) and daily total direct normal irradiance (DNI_sum) (e.g., Figure 2 (As shown in the dark red blocks in the image), therefore they were considered redundant features and were not adopted.

[0068] The final selected and adopted feature subset is a combination with clearly defined physical meaning and low internal correlation. This subset specifically includes: daily total horizontal irradiance (GHI_sum), daily total direct normal irradiance (DNI_sum), daily total diffuse horizontal irradiance (DHI_sum), daily average temperature (Temperature_mean), daily temperature standard deviation (Temperature_std), average cloud cover (Cloud_Opacity_mean), daily average humidity (Humidity_mean), daily average air pressure (Pressure_mean), irradiance centroid time (GHI_shape_energy_centroid_hour), variability index (GHI_shape_variability_diff_std), and irradiance morning ramp-up rate (GHI_shape_ramp_up_avg). This selection step ensures that the feature inputs used for subsequent Gaussian Mixture Model (GMM) clustering are effective and robust.

[0069] S12. Feature Standardization: To eliminate the impact of dimensional differences between daily features on subsequent cluster analysis, Z-score standardization is used. This process involves subtracting the mean and dividing by the standard deviation for each feature dimension. Specifically, for any given day's feature vector... (where the feature dimension D=11), its standardized vector The first in element The calculation is as follows:

[0070] ;

[0071] in, For the standardized first Each feature element; For the original first Each feature element; For the first The mean of each feature across all date samples; For the first The standard deviation of each feature across all date samples.

[0072] S13. Clustering and Pattern Determination: The Gaussian Mixture Model (GMM) clustering algorithm was used to perform cluster analysis on the standardized daily feature matrix. To scientifically determine the optimal number of clusters K, this invention analyzed the information criterion scores under different K values, and the analysis results are as follows: Figure 3 As shown. Figure 3 The results show the score variations of the Bayesian Information Criterion (BIC) and the Akaike Information Criterion (AIC) under different numbers of clusters (K). Figure 3The ordinate represents the information criterion value; a lower value indicates a better model. As shown in the figure, the BIC score reaches its lowest (optimal) when K=4, as indicated by the red dashed line. Therefore, this invention determines the optimal number of clusters, K=4.

[0073] The clustering process automatically groups dates with similar operating characteristics into the same operating pattern. The four patterns identified and named in this embodiment are distributed monthly in 2019 as follows: Figure 4 As shown. Figure 4 The horizontal axis represents the month, and the vertical axis represents the number of days the pattern occurred. As shown in the legend, these four patterns are: Spring / Autumn Medium-Efficiency, Autumn / Winter Medium-Efficiency, Summer Peak, and Low-Efficiency / Cloudy. Their average daily power curves are shown below. Figure 5 As shown. Figure 5 The horizontal axis represents time (hours:minutes), and the vertical axis represents average active power (MW). As shown in the legend, the four modes (and their total number of days in this embodiment) are: 'Inefficient / Cloudy (85 days)', 'Summer Peak (91 days)', 'Spring and Autumn Medium Efficiency (110 days)', and 'Autumn and Winter Medium Efficiency (79 days)'. These four curves show significant differences in peak power, start-stop time, and curve shape, further demonstrating the effectiveness of the mode classification in step one. Finally, the daily mode labels are mapped back to 15-minute data, and a mode affiliation is assigned to each data point.

[0074] It should be noted that the purpose of using daily features for clustering and assigning unique pattern labels to all data throughout the day in this invention is to achieve an effective division of labor and decoupling between 'macro-level pattern recognition' and 'micro-level time series prediction':

[0075] 1. Macroscopic Pattern Recognition: First, at the daily scale, distinguish different types of macroscopic operational backgrounds (e.g., "summer peak day" vs. "inefficient / cloudy day"). These macroscopic backgrounds are reflected in the overall shape of the signal (e.g., Figure 5 Significant differences exist in signal complexity (fuzzy entropy) and key meteorological factors. This granularity of classification ensures the robustness of pattern recognition and avoids interference from intraday instantaneous weather fluctuations on macroscopic operating condition judgments.

[0076] 2. Micro-level Time Series Prediction: The subsequent BiLSTM-Transformer deep learning model is responsible for specifically handling and capturing intraday dynamic changes given the current macro-level background (through pattern ID). Sudden intraday weather changes (such as sunny in the morning and cloudy in the afternoon) are the core micro-level problem that the model needs to solve in step three of this invention. By learning combinations of high-frequency meteorological feature sequences and pattern embedding vectors, the model, as verified by this invention, is fully capable of handling the prediction task of multiple different intraday curves under the same macro-level pattern (such as 'spring and autumn mid-term effects').

[0077] This design enables the model to adaptively invoke prediction logic that matches the current macroscopic model through a mode embedding mechanism, effectively solving the fundamental problems of lack of model awareness and insufficient robustness in the background technology. During the model prediction phase, daily characteristics are calculated based on meteorological data provided by numerical weather prediction, thereby using the trained GMM model to determine the operational mode label for the day to be predicted.

[0078] S2. For each operating mode identified in step S1, independently perform signal decomposition and feature extraction on its corresponding photovoltaic power time series, and calculate signal complexity features (such as fuzzy entropy) based on the decomposition results. Combined with mode-specific external meteorological and temporal features, generate mode-specific input features that match the characteristics of each mode. The specific implementation method is as follows:

[0079] S21. Independent decomposition: Photovoltaic power time series for each operating mode. The improved complete set empirical mode decomposition (ICEEMDAN) method is used independently to decompose it into a series of intrinsic mode functions. and a residual term Its expression is:

[0080] ;

[0081] in, It is a pattern The number of IMFs obtained from the time series decomposition of photovoltaic power. The decomposition yields the first... One Intrinsic Mode Function (IMF) For the residual terms; each intrinsic mode function (IMF) represents the local oscillation characteristics of the signal at different frequency scales;

[0082] The ICEEMDAN decomposition process relies on a set of key hyperparameters, which are among the targets of the subsequent hyperparameter optimization steps in this invention. These parameters specifically include, in this embodiment:

[0083] ce_epsilon (noise standard deviation): controls the amplitude of the auxiliary white noise added during the decomposition process;

[0084] ce_trials (number of integrated trials): refers to the average number of trials performed to obtain a stable IMF component;

[0085] ce_max_imf (maximum number of IMFs): limits the total number of modal functions that can be decomposed.

[0086] The basis for selecting these parameters for optimization in this invention is that they together determine the fineness of ICEEMDAN decomposition (controlled by ce_max_imf) and the noise level (determined by ce_epsilon and ce_trials).

[0087] S22, Adaptive Alignment: To unify the input dimensions of subsequent models, the decomposed... One IMF component and one residual term (total) The components are forcibly recombined into a fixed number of N component sequences (N=9 in this embodiment); the recombination rules are: a) retain the first N-1 (i.e., 8) IMF components. If the actual number of IMFs is less than N-1, then the remaining IMF component positions remain at zero.

[0088] b) Retain one residual term and place it in the Nth component position;

[0089] c) If the actual IMF amount If there are more than N-1 IMF components, the extra (i.e., the Nth and subsequent) IMF components are summed point by point and added to the Nth component (residual term) to ensure dimensionality consistency.

[0090] The advantage of this invention's 'accumulation' strategy (rather than 'discarding' strategy) is that it ensures the integrity of the mathematical information in the signal decomposition (i.e., the sum of the N components can still accurately reconstruct the original signal). More importantly, the feature extraction in this invention is performed only on the first N-1 IMF components; the Nth 'composite residual component' does not participate in subsequent feature calculations. Therefore, this alignment operation ensures information integrity without introducing confusion into the feature set of the input model.

[0091] S23. Signal complexity feature extraction: For each mode, independently calculate the fuzzy entropy value of the N-1 IMF component sequences (excluding residual components) after decomposition and alignment. And these calculated FE values ​​themselves (e.g.) , ,..., It is directly used as a feature to characterize the complexity of the signal and incorporated into the final pattern specialization input feature set.

[0092] The calculation of fuzzy entropy also relies on a set of key hyperparameters, which are also among the targets of the subsequent hyperparameter optimization steps in this invention. These parameters specifically include, in this embodiment:

[0093] fe_m (embedding dimension): The embedding dimension used during phase space reconstruction;

[0094] fe_r_factor (tolerance factor): The tolerance radius used to calculate similarity (usually multiplied by the standard deviation to obtain the specific radius);

[0095] fe_n (fuzzy exponent): the power of a fuzzy function (such as an exponential function);

[0096] fe_tau (time delay): The time delay used during phase space reconstruction.

[0097] The basis for selecting these parameters for optimization in this invention is that they collectively determine the sensitivity of the signal complexity feature (i.e., the fuzzy entropy value).

[0098] S24. External Feature Processing: Define and filter relevant weather feature subsets for each pattern, and extract cyclic coding features based on timestamps. The sine and cosine transforms of hours are obtained by dividing the hour value of a day (e.g., values ​​0-23) by the period 24, multiplying by 2π, and then calculating its sine and cosine values ​​respectively; the sine and cosine transforms of days within the year are obtained by calculating the sine and cosine values ​​of the day of the year ( Divide by the total number of days in that year ( The sine and cosine values ​​of the days of the week are calculated by multiplying the day of the week (365 in a common year and 366 in a leap year) by 2π. The sine and cosine transformations of the days of the week are obtained by dividing the day of the week (e.g., values ​​0-6) by the period 7, multiplying by 2π, and then calculating the sine and cosine values ​​respectively. The calculation formula is as follows:

[0099] ;

[0100] ;

[0101] ;

[0102] in: For time step The hour value (range 0-23); For time step The corresponding day number within the year (i.e., the day of the year); For time step The corresponding day of the week (value range 0-6); This is the total number of days in a year (365 in a common year and 366 in a leap year). These represent the sine and cosine transform results for the corresponding time features, respectively. In addition, lag sequences for multiple past time steps (96 in this embodiment, i.e., 24 hours) are created for the target power and weather features.

[0103] S3. The pattern-specific input features generated in step S2 and the pattern labels assigned in S1 are fed into a pre-defined unified deep learning prediction model that includes a pattern embedding mechanism. The unified deep learning prediction model can adaptively adjust its prediction behavior according to the current operating mode and finally output the predicted value of photovoltaic power generation.

[0104] The unified deep learning prediction model is a pattern-aware BiLSTM-Transformer hybrid model, and its detailed internal structure is as follows: Figure 6 As shown, the model receives two types of features:

[0105] Input characteristics: Pattern ID (0-3 integers).

[0106] Core Features: The core feature is a high-dimensional vector, the specific composition and dimensionality of which depend on the hyperparameter N_STEPS_IN (input time step, 96 in this example) and the current operating mode. Specifically, it includes: 1. Fuzzy Entropy Feature: 8 floating-point numbers (corresponding to the fuzzy entropy values ​​of N-1=8 IMF components). 2. Time Feature: 6 floating-point numbers (Sin / Cos cyclic encoding of hour, day of the year, and day of the week). 3. Current Weather Feature: Dimensional variable (e.g., the "Spring and Autumn Medium-Effect" mode uses 5 weather features (GHI, DNI, Temperature, Cloud_Opacity, Humidity)). 4. Lag Power Feature: N_STEPS_IN features (e.g., normalized power values ​​from the past 96 time steps). 5. Lag Weather Feature: N_STEPS_IN * (number of current weather features) features (e.g., 96 steps * 5 weather features = 480 features). Therefore, for the model "Spring and Autumn Mid-Effect" (5 weather features, N_STEPS_IN=96), the total dimensions of the "core features" are 8+6+5+96+480=595.

[0107] The pattern ID is fed into the pattern embedding layer, and its output embedding vector is concatenated with the core features in the feature fusion module to form a fused feature vector.

[0108] The fused feature vector is then passed through a first-layer bidirectional LSTM and a second-layer bidirectional LSTM to capture temporal dependencies.

[0109] The output sequence of the second-layer bidirectional LSTM has two data flow paths: 1. Main path: The sequence is fed into the context weighter (Transformer self-attention) module to capture long-range dependencies. 2. Residual path: The sequence bypasses the Transformer module through residual connections.

[0110] The output of the Transformer module (main path) and the output of the residual connection (residual path) are added together and then sent to the layer normalization module for processing.

[0111] Finally, the vector after layer normalization is fed into the final output layer to obtain the output: scaled active power (1 floating-point number).

[0112] In one specific embodiment, the unified deep learning prediction model is a hybrid model architecture, such as... Figure 6 As shown, the model is built using the PyTorch framework, and its structure and data flow include the following:

[0113] Pattern Embedding Layer: To achieve the pattern awareness capability of this invention, an embedding layer (torch.nn.Embedding) is set at the front end of the model. For the input data at each time step, its corresponding pattern label mt (integer 0-3) is fed into the embedding layer to generate a low-dimensional, dense pattern embedding vector that can be learned during model training. :

[0114] ;

[0115] ;

[0116] in: In time step The pattern ID (an integer, such as 0, 1, 2, or 3). The lookup function for the embedding layer (i.e.) It retrieves the corresponding row vector from the embedding matrix based on the input integer ID. It is the output pattern embedding vector, which is a... Dense vectors in space. It is the preset embedding dimension, a hyperparameter that determines the richness of the pattern representation.

[0117] Feature fusion module: the pattern embedding vector Compared with other pattern-specific input feature vectors at time t (Right now Figure 6 The core features shown are concatenated (i.e., feature fusion) to form an enhanced feature vector. This vector will serve as the unified input for subsequent network layers:

[0118] ;

[0119] in: In time step The enhanced final input vector. Represents the time step The vector consisting of all the original input features has a dimension of . .

[0120] Bidirectional Long Short-Term Memory (BiLSTM) layer: the enhanced feature vector sequence It is fed into two cascaded BiLSTM layers. This is used to capture the local contextual dependencies and temporal information of the input sequence.

[0121] Multi-head self-attention (Transformer) layer: This layer is set after the BiLSTM layer and is used to receive the output of the BiLSTM layer. It efficiently captures long-term dependency patterns through the self-attention mechanism, that is, it captures long-range dependencies in the sequence.

[0122] Through this "embedding-splitting-processing" mechanism, pattern information (in...) The embedding vector (in the form of a vector) is explicitly injected into the network. Within the framework of independent training, this embedding vector acts as a learnable contextual prior, providing baseline feature distribution information about that specific operating mode (such as 'summer peak') for the weight updates of the current dedicated model, thereby accelerating model convergence and improving fitting accuracy.

[0123] The training and prediction of a unified deep learning prediction model are explained in detail below:

[0124] First, regarding the training objective, the training objective of the model is to minimize the mean squared error (MSE) loss between the normalized power prediction and the normalized true power value of the model output.

[0125] The model's structure, regularization, and training process also rely on a set of key hyperparameters, which are also among the targets of the subsequent hyperparameter optimization steps in this invention. These parameters specifically include, in this embodiment:

[0126] Model structure parameters include, for example, lstm1 and lstm2 (representing the number of hidden units in the first and second BiLSTM layers, respectively) and heads (representing the number of multi-head self-attention heads in the Transformer layer). These are selected for optimization to match the sequence complexity of different patterns.

[0127] Regularization parameters include dropout (representing the random deactivation rate applied between network layers) and l2 (representing the weight coefficients of the L2 regularization term). These are chosen for optimization to prevent overfitting of the model across different datasets.

[0128] Training dynamic parameters include learning rate and batch size. These are chosen for optimization to ensure effective convergence for each model.

[0129] The formula for calculating the training objective is:

[0130] ;

[0131] in: This represents the total loss function value. This represents the number of samples in the training batch (Batch Size). For the first Normalized true power values ​​for each sample; The first output of the model Normalized power prediction for each sample; is the L2 regularization coefficient (corresponding to the hyperparameter l2 in the embodiment); These are all trainable weight parameters in the model; This represents the square of the L2 norm.

[0132] Secondly, regarding the acquisition of the final power prediction, this process is achieved through a multi-stage process: the first stage is to perform inverse scaling: by performing an inverse scaling operation on the normalized power prediction value directly output by the model, an intermediate power prediction value is obtained; the second stage is physical constraints: then, a two-stage physical constraint process based on physical scene priority is applied to the intermediate power prediction value.

[0133] In a preferred embodiment, the two-stage physical constraint processing steps include:

[0134] S31. Nighttime Constraints: First, obtain the solar zenith angle corresponding to the predicted time point. The solar zenith angle, along with GHI and temperature, is an external feature included in the original photovoltaic operation data.

[0135] S32. Physical zeroing: Based on the obtained zenith angle, determine whether the solar zenith angle is greater than or equal to 90 degrees (i.e., it is nighttime). If so, force the final power prediction value for that time point to be set to 0.

[0136] S33. Non-negativity constraint: If the judgment result is negative (i.e., it is determined to be daytime), it indicates that it is not nighttime. Therefore, a non-negativity constraint should be applied to the intermediate power prediction value, i.e., the maximum value between the intermediate power prediction value and 0 should be taken as the final power prediction value for that time point. Its implementation can be expressed by the following formula:

[0137] ;

[0138] This represents the final power prediction value; This represents the intermediate power prediction value. The two-stage processing (i.e., using MSE during training and applying physical constraints during prediction) employed in this invention for both the training objective and the final power prediction ensures a design that unifies mathematical optimality with physical reality.

[0139] To maximize the performance of the entire framework, this invention also includes a mode-specific hyperparameter optimization step. In this embodiment, the Optuna framework based on Bayesian optimization is used to independently search and optimize the ICEEMDAN decomposition parameters, fuzzy entropy calculation parameters, and BiLSTM-Transformer model structure and training parameters for four identified operating modes, in order to find the parameter combination with the lowest root mean square error (RMSE) on each mode validation set.

[0140] The parameter space optimized by the hyperparameter optimization step of this invention covers the three key stages of this invention. The specific parameters in this space (such as ce_epsilon, ce_trials, ce_max_imf, fe_m, fe_r_factor, fe_n, fe_tau, lstm1, lstm2, heads, dropout, l2, learning_rate, batch_size, etc.) and their selection criteria have been described in detail in their respective technical steps.

[0141] In this embodiment, the optimal parameter combinations determined for the four modes through the above hyperparameter optimization steps are shown in the following examples:

[0142]

[0143] In the final evaluation, this invention does not construct completely different independent models for each mode, but instead adopts a more efficient strategy of "same architecture, parameter specialization, independent training":

[0144] Same architecture: All prediction tasks use a unified BiLSTM-Transformer model architecture;

[0145] Parameter specialization: For each operating mode (e.g., 'summer peak'), load the optimal hyperparameters specific to that mode (e.g., different numbers of lstm1 units, learning_rate, etc.) from the configuration dictionary obtained from the previously described mode specialization hyperparameter optimization steps.

[0146] Independent training: In each Temporal Cross-Validation Fold (TSCV-Fold) of the final evaluation, the model is independently retrained as a completely new model instance for the data of that pattern (train_new_flag=True).

[0147] This strategy ensures the consistency of the model architecture, while maximizing the adaptive fitting of the dynamic characteristics of each mode through hyperparameter specialization and independent training. Its generalization ability is also rigorously evaluated through TSCV, effectively avoiding overfitting.

[0148] In summary, through the coordinated operation of the above steps, this invention successfully realizes a complete framework capable of automatically identifying photovoltaic operating modes and performing adaptive high-precision predictions, solving the "mode agnostic" problem existing in the prior art. The final prediction and evaluation results are as follows: Figure 7 and Figure 8 As shown. The results were obtained on a separate final test set. Figure 7 This is a comparison chart of the independent evaluation results for each mode. The chart contains four sub-charts, each corresponding to one of the four operating modes identified by this invention. Each sub-chart compares the actual power (solid black line) with the predicted power (dashed red line). Figure 7 As shown, in all four modes of the subplots, the red dashed line representing the predicted power closely matches and overlaps with the black solid line representing the actual power, which intuitively proves that the present invention achieves high-precision prediction in all modes.

[0149] To further quantify this conclusion, specific evaluation metrics (root mean square error RMSE, mean absolute error MAE, coefficient of determination R) are used. 2 This is also shown in each sub-graph and summarized below:

[0150] Inefficient / Cloudy Mode: RMSE=2.973, MAE=1.439, R 2 =0.9756;

[0151] Summer peak season mode: RMSE=4.412, MAE=2.672, R 2 =0.9859;

[0152] Spring and Autumn Medium-Efficiency Mode: RMSE=4.535, MAE=2.452, R 2 =0.9743;

[0153] Autumn / Winter Medium-Efficiency Mode: RMSE=1.726, MAE=0.823, R 2 =0.9952.

[0154] Figure 8This is a scatter plot of the predicted vs. actual values ​​for the entire year. The horizontal axis, ActualPower (MW), represents the actual power, and the vertical axis, PredictedPower (MW), represents the predicted power. As shown, the scatter plots (Predicted vs. Actual) for all evaluation time points are closely distributed around the ideal diagonal of y=x indicated by the red dashed line, further demonstrating the high accuracy and robustness of this invention under different operating conditions; and exhibiting significant superiority in handling the seasonal variations and complex dynamic characteristics of photovoltaic power generation.

[0155] Within the technical scope disclosed in this invention, any variations or substitutions that can be easily conceived should be included within the protection scope of this invention. Therefore, the protection scope of this invention should be determined by the scope of the claims.

Claims

1. A photovoltaic power generation power prediction method based on pattern perception adaptive deep learning, characterized in that, The method comprises the following steps: S1, based on the historical operation data of the photovoltaic power station, extracting the daily insolation feature and automatically identifying and dividing at least two operation modes, and mapping the identified operation mode label to all data time points within the natural day corresponding to the daily insolation feature; S2, for each operation mode identified in step S1, independently performing signal decomposition and feature extraction on the corresponding photovoltaic power time sequence, and calculating the signal complexity feature based on the decomposition result, combining the mode-specific external meteorological and time features to construct a feature set adapted to each mode; The independently performing signal decomposition and feature extraction on the corresponding photovoltaic power time sequence comprises: S21, independent decomposition: independently performing ICEEMDAN decomposition on the photovoltaic power time sequence under each operation mode, and decomposing the original photovoltaic power time sequence into the sum of multiple intrinsic mode functions (IMF) and a residual term; S22, adaptive alignment: to unify the input dimension of the subsequent model, the decomposed components are recombined into a fixed number N of component sequences; the rule for recombining the decomposed components into a fixed number N of component sequences is: a) retain the first N-1 IMF components, if the actual number of IMFs is less than N-1, the remaining IMF component positions are kept as zero; b) retain one residual term and place it in the Nth component position; c) if the actual number of IMFs is more than N-1, sum the extra IMF components point by point and accumulate them on the Nth component to ensure dimension uniformity; S23, signal complexity feature extraction: for the IMF component sequence decomposed and aligned under each mode, independently calculate the fuzzy entropy value, and include these fuzzy entropy values themselves as signal complexity features into the feature set; S3, based on a bidirectional long short-term memory network, a unified deep learning prediction model containing a mode embedding layer is constructed, for each operation mode identified in step S1, independently load the optimized exclusive parameters thereof, and input the feature set adapted to the operation mode constructed in step S2 and the mode label into the unified deep learning prediction model for training, and output the power prediction result. 2.The photovoltaic power generation power prediction method based on mode perception adaptive deep learning according to claim 1, wherein, The step S2 further comprises generating mode-specific external features for each operation mode, and extracting timestamp-based cyclic encoding features; the cyclic encoding features include sine and cosine transformations of hours, days within a year, and weekdays. 3.The photovoltaic power generation power prediction method based on mode perception adaptive deep learning according to claim 1, characterized in that, The step S2 further comprises generating mode-specific external features for each operation mode, and extracting timestamp-based cyclic encoding features; the cyclic encoding features include sine and cosine transformations of hours, days within a year, and weekdays. The step S2 further comprises generating mode-specific external features for each operation mode, and extracting timestamp-based cyclic encoding features; the cyclic encoding features include sine and cosine transformations of hours, days within a year, and weekdays. ​ S13, clustering and mode determination: the normalized daily profile features are clustered by using Gaussian Mixture Model clustering algorithm, and the optimal K value is selected as the best mode number by calculating the Bayesian Information Criterion BIC under different K values, and the dates with similar features are classified into the same cluster, and each cluster is defined as a running mode.

4. The photovoltaic power generation power prediction method based on mode perception adaptive deep learning according to claim 3, characterized in that, The Z-score standardization method is used for processing, including: for any feature element in the daily profile feature vector of any day, subtracting the mean value of the feature in all date samples, and dividing by the standard deviation of the feature in all date samples, to obtain the standardized feature element. 5.The photovoltaic power generation power prediction method based on mode perception adaptive deep learning according to claim 1, wherein, The unified deep learning prediction model is a hybrid model architecture, and the structure thereof comprises: a mode embedding layer for receiving the mode label assigned in step S1 and mapping it into a mode embedding vector; a feature fusion module for splicing the mode embedding vector and the feature set generated in step S2 to form a fusion feature vector; a bidirectional long short-term memory network layer for receiving the fusion feature vector to capture local context dependency of the input sequence; a multi-head self-attention layer for receiving the sequence containing local context dependency output by the bidirectional long short-term memory network layer to capture long-range dependency in the input sequence. 6.The photovoltaic power generation power prediction method based on pattern perception adaptive deep learning according to claim 5, wherein, The unified deep learning prediction model comprises: training target: the training target of the model is to minimize the mean square error loss between the normalized power prediction value output by the model and the normalized true power value; final power prediction: an intermediate power prediction value is obtained by performing reverse scaling operation on the normalized power prediction value directly output by the model; then, two-stage physical constraint processing based on physical scene priority is applied to the intermediate power prediction value to obtain the final photovoltaic power prediction value. 7.The photovoltaic power generation power prediction method based on pattern perception adaptive deep learning according to claim 6, characterized in that, The final power prediction comprises: S31, night constraint: obtaining the solar zenith angle corresponding to the prediction time point; S32, physical zeroing: judging whether the solar zenith angle is greater than or equal to 90 degrees, if yes, the final power prediction value of the time point is forcibly set to 0; S33, non-negative constraint: if not, the non-negative constraint is applied to the intermediate power prediction value, and the maximum value between the intermediate power prediction value and 0 is taken as the final power prediction value of the time point.

Images