Photovoltaic power prediction analysis method and device based on two-stage hybrid deep learning

Abstract

The application discloses a photovoltaic power prediction analysis method and device based on two-stage hybrid deep learning, and belongs to the photovoltaic power generation field. The method aims at the problems that a deep learning model is prone to overfitting under weather changes, and a traditional mapping method cannot capture power fluctuation details, and constructs a two-stage hybrid deep learning framework. In the first stage, a mapping model based on a long short-term memory network is used to input numerical weather forecast data, and the overall trend of power is output. In the second stage, a wavelet decomposition combined with a long short-term memory network model is used to extract adjacent day time series features, and a convolutional neural network model is used to extract similar day weather pattern features, so that fine fluctuation information is obtained. Finally, a time segmentation integration method is used to dynamically weight and fuse three results, and grid search and Bayesian optimization are combined to realize automatic optimization of hyperparameters. The application considers both prediction accuracy and adaptability, and is excellent in interval day-ahead prediction among photovoltaic power stations of different scales, and can accurately predict power fluctuation details under different weathers.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power generation technology, specifically relating to a photovoltaic power prediction and analysis method and device based on two-stage hybrid deep learning. Background Technology

[0002] Photovoltaic power generation is a core component of the clean energy system. The output of photovoltaic power plants exhibits significant randomness and volatility, posing serious challenges to the stable operation of the power grid, power dispatch planning, and electricity market transactions. High-precision day-ahead photovoltaic power forecasting can effectively reduce the impact of photovoltaic grid connection on the power grid, and is crucial for ensuring grid security and stability, optimizing the allocation of generation-side resources, and improving the economic benefits of photovoltaic power plants.

[0003] Existing photovoltaic power prediction methods are mainly divided into three categories. The first category is physical methods, which are based on the photoelectric conversion principle and rely heavily on the accuracy of equipment parameters and environmental monitoring, and have poor adaptability to complex weather conditions. The second category is statistical methods, which establish mapping relationships through historical data, but have insufficient nonlinear fitting capabilities and are difficult to capture the details of power fluctuations. The third category is deep learning methods, represented by models such as Long Short-Term Memory Networks (LSTM) and Convolutional Neural Networks (CNN), which have significant advantages in feature extraction. Related research aims to improve prediction performance by improving single models or constructing hybrid frameworks.

[0004] However, existing technologies share common drawbacks. Deep learning models are prone to overfitting and have weak generalization ability when inter-day weather conditions change significantly. Mapping methods based on numerical weather prediction (NWP) are stable but cannot accurately predict the fine details of power fluctuations. The difficulty in balancing prediction accuracy and model adaptability has become a pressing issue in the field of day-ahead photovoltaic power prediction. Summary of the Invention

[0005] Therefore, the present invention provides a photovoltaic power prediction and analysis method and apparatus based on two-stage hybrid deep learning to solve or partially solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: Firstly, a photovoltaic power prediction and analysis method based on two-stage hybrid deep learning is provided, comprising the following steps:

[0007] Preprocessing and error correction are performed on the acquired raw numerical weather forecast data to obtain corrected numerical weather forecast data; A two-stage hybrid deep learning prediction model framework is constructed, which includes an overall trend prediction stage based on numerical weather prediction mapping and a fluctuation prediction stage based on historical feature extraction. In the overall trend prediction stage of the numerical weather forecast mapping, a mapping model based on a long short-term memory network is established and trained. Using historical measured meteorological data and corresponding historical photovoltaic power data as training sets, the mapping relationship between meteorological conditions and power generation is learned. The corrected numerical weather forecast data is input, and the overall trend prediction result of photovoltaic power is output. In the fluctuation prediction stage of historical feature extraction, dual-path historical feature extraction is performed in parallel. The first path uses a wavelet decomposition combined with a long short-term memory network model to decompose the historical photovoltaic power sequence and extract the temporal features of adjacent days, outputting the first path fluctuation prediction result. The second path uses a convolutional neural network model to extract similar day weather pattern features based on historical data of weather type clustering, outputting the second path fluctuation prediction result. A time-segmented integration method is adopted, and the prediction interval is divided according to the rising, fluctuating and falling periods of the photovoltaic power generation daily power curve. Dynamic weights are assigned to the overall trend prediction result, the first path fluctuation prediction result and the second path fluctuation prediction result and then weighted and integrated. An algorithm combining grid search and Bayesian optimization is used to automatically optimize the hyperparameters of the mapping model based on long short-term memory network, the wavelet decomposition combined with long short-term memory network model, and the convolutional neural network model. The robustness of the model is evaluated by K-fold cross-validation, and the final day-ahead photovoltaic power prediction result is output.

[0008] As a preferred scheme for photovoltaic power prediction and analysis based on two-stage hybrid deep learning, the process of preprocessing and error correction of the acquired raw numerical weather prediction data to obtain corrected numerical weather prediction data specifically includes: Interpolation methods are used to interpolate the raw numerical weather forecast data to match the time resolution of the photovoltaic power data; Based on historical measured meteorological datasets and historical numerical weather prediction datasets, a time-by-time error correction method is used to reduce prediction errors. The time-by-time error correction process satisfies the following formula:

[0009] In the formula, For corrected numerical weather forecast data, The original numerical weather forecast data to be corrected. This is the error correction value. For the first Heavenly The error value between numerical weather forecast data and measured data at a given time point For the first Error weighting for each day It is a set of days for historical data. For the first time in history Heavenly Numerical weather forecast data at specific points in time, For the first time in history Heavenly Measured meteorological data at specific time points; The error correction value Let be the set of correction values ​​for solar irradiance, relative humidity, and temperature, satisfying the formula:

[0010] In the formula, This is the error correction value for solar irradiance. This is the error correction value for relative humidity. This is the error correction value for temperature.

[0011] As a preferred scheme for photovoltaic power prediction and analysis based on two-stage hybrid deep learning, a mapping model based on a long short-term memory network is established and trained. Using historical measured meteorological data and corresponding historical photovoltaic power data as the training set, the mapping relationship between meteorological conditions and power generation is learned. During the training and prediction of this mapping relationship, the formula is satisfied:

[0012]

[0013] In the formula, This refers to the mapping relationship during the model training phase. This represents the mapping relationship during the model testing phase. For the hyperparameters of the mapping model, , , These are historically measured data for solar irradiance, relative humidity, and temperature. Historical photovoltaic power data, , , These are the predicted solar irradiance, relative humidity, and temperature data for the forecast day. This is the overall trend prediction result.

[0014] As a preferred scheme for photovoltaic power prediction and analysis based on two-stage hybrid deep learning, the first path utilizes a wavelet decomposition combined with a long short-term memory network model to decompose historical photovoltaic power sequences and extract the temporal features of nearby days. During the output of the first path fluctuation prediction results, the wavelet decomposition combined with the long short-term memory network model calculates the following formula:

[0015]

[0016]

[0017]

[0018]

[0019]

[0020] In the formula, , , , , , , This is the weight matrix. , , , For bias vectors, Forget gate activation vector, It is the Sigmoid activation function. Input to Long Short-Term Memory (LSTM) network units, The input gate activation vector, To update the gate output value, It is the ReLU activation function. For scalar multiplication, For the unit state, The output gate activation vector, This is the output of the Long Short-Term Memory (LSTM) network unit.

[0021] As a preferred approach for photovoltaic power prediction and analysis based on two-stage hybrid deep learning, the extraction of near-day time-series features after decomposing historical photovoltaic power sequences specifically includes: The historical photovoltaic power sequence is decomposed into subsequences of different frequencies using wavelet decomposition, which satisfies the following formula:

[0022]

[0023] In the formula, It is a low-frequency sequence. , It is a high-frequency sequence. As a scale factor, The translation factor is... This is a historical photovoltaic power sequence. For wavelet basis functions, yes The conjugate function of the decomposed sequence is used; the long short-term memory network model is used to predict the decomposed sequence respectively, and the predicted values ​​are summed to obtain the fine fluctuation prediction results corresponding to the time series characteristics of the nearby days.

[0024] As a preferred scheme for photovoltaic power prediction and analysis based on two-stage hybrid deep learning, the first path utilizes a model combining wavelet decomposition and long short-term memory network to decompose the historical photovoltaic power sequence and extract the temporal features of the nearest day. In the process of outputting the first path fluctuation prediction result, the first path fluctuation prediction result corresponding to the temporal features of the nearest day satisfies the following formula:

[0025] In the formula, The prediction results are for the time series features of nearby days. For the time series model of Long Short-Term Memory Network, For model hyperparameters, This is the decomposed historical power subsequence.

[0026] As a preferred scheme for photovoltaic power prediction and analysis based on two-stage hybrid deep learning, the second path utilizes a convolutional neural network model to extract similar daily weather pattern features from historical data clustered by weather type, and outputs the second path fluctuation prediction results: Historical photovoltaic power data is clustered and categorized according to weather type. Based on the weather type of the predicted day, historical data of the corresponding category are selected and input into a convolutional neural network model for training and prediction. This yields the second path fluctuation prediction results corresponding to the weather pattern characteristics of similar days, satisfying the formula:

[0027] In the formula, To predict the forecast results corresponding to the daily weather type, It is a convolutional neural network model. This data represents historical power levels for the same weather type. These are the hyperparameters of the model for the corresponding weather type. To predict the daily weather type, Historical weather types.

[0028] As a preferred scheme for photovoltaic power prediction and analysis methods based on two-stage hybrid deep learning, the information extraction process of the convolutional neural network satisfies the following formula:

[0029] In the formula, For the first Layer Each convolutional kernel is located at... The output, For the first Layer The weights of each convolutional kernel, For the first Layer position The input area, For the first Layer The bias of each convolution kernel This is the activation function.

[0030] As a preferred scheme for photovoltaic power prediction and analysis based on two-stage hybrid deep learning, a time-segmented integration method is adopted. The prediction interval is divided according to the rising, fluctuating, and falling periods of the daily photovoltaic power generation curve. Dynamic weights are assigned to the overall trend prediction result, the first path fluctuation prediction result, and the second path fluctuation prediction result, and during the weighted fusion process, the time-segmented integration method satisfies the following formula:

[0031]

[0032]

[0033] In the formula, For the first The final predicted value for each time period, , , These are the overall trend model, the nearby day time series feature model, and the similar day weather pattern feature model, respectively, at the [number]th [year]. Weighting coefficients for each time period , , The corresponding models are in the 1st month. Predicted values ​​for a given time period This represents the final overall prediction result.

[0034] As a preferred scheme for photovoltaic power prediction and analysis based on two-stage hybrid deep learning, an algorithm combining grid search and Bayesian optimization is used to automatically optimize the hyperparameters of the mapping model based on a long short-term memory network, the wavelet decomposition combined with a long short-term memory network model, and the convolutional neural network model. The robustness of the model is evaluated through K-fold cross-validation, and the final day-ahead photovoltaic power prediction result is output. The optimal value of the hyperparameters obtained through Bayesian optimization satisfies the following formula:

[0035] In the formula, The optimal value for the hyperparameter is... For the hyperparameters to be optimized, For the search space of hyperparameters, The objective function is denoted as .

[0036] Secondly, the present invention provides a photovoltaic power day-ahead prediction device based on two-stage hybrid deep learning, employing the photovoltaic power prediction and analysis method based on two-stage hybrid deep learning as described in the first aspect or any possible implementation thereof, including: The data preprocessing module is used to perform preprocessing and error correction on the acquired raw numerical weather forecast data to obtain corrected numerical weather forecast data. A two-stage prediction model construction module is used to construct a two-stage hybrid deep learning prediction model framework, which includes an overall trend prediction stage based on numerical weather forecast mapping and a fluctuation prediction stage based on historical feature extraction. The first-stage processing module is used in the overall trend prediction stage of the numerical weather forecast mapping to establish and train a mapping model based on a long short-term memory network. It uses historical measured meteorological data and corresponding historical photovoltaic power data as training sets to learn the mapping relationship between meteorological conditions and power generation. It inputs the corrected numerical weather forecast data and outputs the overall trend prediction result of photovoltaic power. The second-stage processing module, used in the fluctuation prediction stage of historical feature extraction, performs parallel dual-path historical feature extraction. The first path uses a wavelet decomposition combined with a long short-term memory network model to decompose the historical photovoltaic power sequence and extract the temporal features of adjacent days, outputting the first path fluctuation prediction result. The second path uses a convolutional neural network model to extract similar day weather pattern features based on historical data of weather type clustering, outputting the second path fluctuation prediction result. The weighted fusion module is used to divide the prediction interval according to the rising, fluctuating and falling periods of the photovoltaic power generation daily power curve using a time segmented integration method, and to assign dynamic weights to the overall trend prediction result, the first path fluctuation prediction result and the second path fluctuation prediction result and then weighted fusion them. The hyperparameter optimization and result output module is used to automatically optimize the hyperparameters of the mapping model based on long short-term memory network, the wavelet decomposition combined with long short-term memory network model, and the convolutional neural network model based on an algorithm combining grid search and Bayesian optimization. The module evaluates the robustness of the model through K-fold cross-validation and outputs the final day-ahead photovoltaic power prediction result.

[0037] Thirdly, the present invention provides an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the processor, when executing the program or instructions, implements the photovoltaic power prediction and analysis method based on two-stage hybrid deep learning, as described in the first aspect or any possible implementation thereof.

[0038] Fourthly, the present invention provides a computer-readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of the photovoltaic power prediction and analysis method based on two-stage hybrid deep learning, as described in the first aspect or any possible implementation thereof.

[0039] The present invention has the following advantages: First, the first stage of this invention uses a mapping model based on a long short-term memory network to output the overall trend of power from numerical weather forecast data, ensuring strong adaptability of the model when inter-day weather conditions change significantly and effectively avoiding overfitting. The second stage utilizes wavelet decomposition combined with a long short-term memory network model and a convolutional neural network model to extract temporal features of neighboring days and weather pattern features of similar days in parallel, capturing fine details of power fluctuations and achieving a balance between prediction accuracy and generalization ability.

[0040] Secondly, this invention employs a time-segmented integration method, dynamically allocating the weights of each sub-model according to the rising, fluctuating, and falling periods of the photovoltaic power generation curve, thereby enabling targeted optimization of prediction performance for different time periods. This method performs excellently in high-resolution day-ahead predictions at 15-minute intervals, meeting the grid dispatching requirements for refined power prediction.

[0041] Third, this invention combines a hyperparameter optimization strategy with grid search and Bayesian optimization. Grid search quickly determines hyperparameters such as optimizers and kernel initializers, while Bayesian optimization efficiently solves for the optimal combination of numerical hyperparameters such as training cycle and learning rate. Combined with K-fold cross-validation to evaluate model robustness, this significantly improves model tuning efficiency and final performance.

[0042] Fourth, the method of this invention maintains high accuracy and strong adaptability in actual tests on both 10MW large-scale photovoltaic power plants and 100kW small-scale photovoltaic power plants, and is not limited by the size of the power plant. At the same time, the method is compatible with various weather types such as sunny, cloudy, rainy, and overcast days, and can be widely applied to power prediction scenarios of photovoltaic power plants of different regions and sizes.

[0043] Fifth, this invention can provide a reliable basis for grid dispatching, effectively reducing the impact of the randomness and volatility of photovoltaic output on the stable operation of the grid. This method helps optimize power generation planning, improves the grid connection and absorption capacity of photovoltaic power plants, and thus reduces grid peak-shaving costs and photovoltaic curtailment rates. Attached Figure Description

[0044] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0045] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0046] Figure 1 This is a schematic diagram of the photovoltaic power prediction and analysis method based on two-stage hybrid deep learning provided in an embodiment of the present invention; Figure 2 This is a flowchart of the two-stage prediction process in the photovoltaic power prediction and analysis method based on two-stage hybrid deep learning provided in this embodiment of the invention. Figure 3 This is a schematic diagram of the photovoltaic power prediction and analysis device architecture based on two-stage hybrid deep learning provided in an embodiment of the present invention. Figure 4 This is an electronic device architecture diagram provided in an embodiment of the present invention. Detailed Implementation

[0047] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Photovoltaic power generation is a core component of the clean energy system. The output of photovoltaic power plants exhibits significant randomness and volatility, posing serious challenges to the stable operation of the power grid, power dispatch planning, and electricity market transactions. High-precision day-ahead photovoltaic power forecasting can effectively reduce the impact of photovoltaic grid connection on the power grid, and is crucial for ensuring grid security and stability, optimizing the allocation of generation-side resources, and improving the economic benefits of photovoltaic power plants.

[0049] Currently, photovoltaic power prediction methods are mainly divided into three categories: physical methods, statistical methods, and artificial intelligence-based methods.

[0050] Physical method: This method is based on the principle of photoelectric conversion, combining photovoltaic module parameters, power plant geographical information, and meteorological data for modeling and calculation. The prediction accuracy of the physical method is highly dependent on the accuracy of equipment parameters and environmental monitoring data. Under complex and variable weather conditions, the model has poor adaptability and struggles to accurately capture nonlinear power fluctuations.

[0051] Statistical method: This method constructs a statistical mapping model to achieve prediction by analyzing the correlation between historical power data and historical meteorological data. The statistical modeling process is relatively simple, but the model's nonlinear fitting ability is limited, and it cannot fully explore the complex features hidden in the data, showing significant shortcomings in predicting the details of power fluctuations.

[0052] Artificial intelligence-based approaches: In recent years, deep learning technologies, represented by Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs), have been widely applied in the field of photovoltaic power prediction due to their powerful feature extraction and nonlinear fitting capabilities. Related research has effectively improved the predictive performance of single models in specific scenarios by improving their structures, such as optimizing LSTM models by combining weather classification strategies and using wavelet decomposition techniques to preprocess data before inputting it into the LSTM model. Furthermore, hybrid model frameworks have also become a research hotspot, such as combining CNNs and LSTMs to achieve the collaborative extraction of spatial and temporal features.

[0053] However, existing deep learning-based photovoltaic power prediction methods still suffer from the common problem of struggling to balance prediction accuracy and model generalization adaptability. On the one hand, deep learning models that focus on mining complex features of historical data are prone to overfitting when diurnal weather conditions change significantly, leading to a substantial decrease in the model's robustness and adaptability. On the other hand, the mapping method based on numerical weather prediction (NWP) widely used in power plants, while exhibiting strong stability and good generalization ability, can only predict the overall trend of power changes and cannot accurately depict the fine details of power fluctuations.

[0054] In summary, improving the prediction accuracy of photovoltaic power fluctuation details while ensuring the model's adaptability to different weather conditions has become a key technical problem that urgently needs to be solved in the field of day-ahead photovoltaic power prediction. In view of this, this invention provides a photovoltaic power prediction and analysis method and apparatus based on two-stage hybrid deep learning, which improves the prediction accuracy of photovoltaic power fluctuation details while ensuring the model's adaptability to different weather conditions. The specific details of this invention are as follows.

[0055] See Figure 1 and Figure 2 This invention provides a photovoltaic power prediction and analysis method based on two-stage hybrid deep learning, comprising the following steps: S1. Perform preprocessing and error correction on the acquired raw numerical weather forecast data to obtain corrected numerical weather forecast data; S2. Construct a two-stage hybrid deep learning prediction model framework, which includes an overall trend prediction stage of numerical weather forecast mapping and a fluctuation prediction stage of historical feature extraction. In the overall trend prediction stage of the numerical weather forecast mapping, a mapping model based on a long short-term memory network is established and trained. Using historical measured meteorological data and corresponding historical photovoltaic power data as training sets, the mapping relationship between meteorological conditions and power generation is learned. The corrected numerical weather forecast data is input, and the overall trend prediction result of photovoltaic power is output. In the fluctuation prediction stage of historical feature extraction, dual-path historical feature extraction is performed in parallel. The first path uses a wavelet decomposition combined with a long short-term memory network model to decompose the historical photovoltaic power sequence and extract the temporal features of adjacent days, outputting the first path fluctuation prediction result. The second path uses a convolutional neural network model to extract similar day weather pattern features based on historical data of weather type clustering, outputting the second path fluctuation prediction result. S3. Using a time-segmented integration method, the prediction interval is divided according to the rising, fluctuating, and falling periods of the photovoltaic power generation daily power curve. Dynamic weights are assigned to the overall trend prediction result, the first path fluctuation prediction result, and the second path fluctuation prediction result, and then weighted and integrated. S4. Based on an algorithm combining grid search and Bayesian optimization, the hyperparameters of the mapping model based on long short-term memory network, the wavelet decomposition combined with long short-term memory network model, and the convolutional neural network model are automatically optimized. The robustness of the model is evaluated by K-fold cross-validation, and the final day-ahead photovoltaic power prediction result is output.

[0056] In this embodiment, step S1, which involves preprocessing and error correction of the acquired raw numerical weather forecast data to obtain corrected numerical weather forecast data, specifically includes: Interpolation methods were used to interpolate the raw numerical weather prediction (NWP) data to match the temporal resolution of the photovoltaic (PV) power data. NWP data is a crucial input for day-ahead PV power forecasting. However, raw NWP data typically suffers from low temporal resolution, data gaps, and systematic deviations from local measured values ​​at power plants, necessitating preprocessing and correction.

[0057] First, the raw NWP data is interpolated to match the time resolution of the photovoltaic power data. Since the acquired NWP output has a 1-hour resolution and contains intervals, while the power data has a 15-minute resolution, polynomial interpolation and sample-and-hold interpolation methods are used to obtain interval-free 15-minute resolution NWP data.

[0058] Then, error correction is performed on the NWP data to reduce prediction errors caused by spatial resolution limitations (typically a 10km × 10km grid). This is based on historical measured meteorological datasets. and historical NWP dataset A time-by-time error correction method is employed to improve the NWP data used for forecasting the following day. The quality of the data is improved. Specifically, based on historical measured meteorological datasets and historical numerical weather prediction datasets, a time-by-time error correction method is used to reduce prediction errors. This time-by-time error correction process satisfies the following formula:

[0059] In the formula, For corrected numerical weather forecast data, The original numerical weather forecast data to be corrected. This is the error correction value. For the first Heavenly The error value between numerical weather forecast data and measured data at a given time point For the first Error weighting for each day It is a set of days for historical data. For the first time in history Heavenly Numerical weather forecast data at specific points in time, For the first time in history Heavenly Measured meteorological data at specific time points; Wherein, the error correction value Let be the set of correction values ​​for solar irradiance, relative humidity, and temperature, satisfying the formula:

[0060] In the formula, This is the error correction value for solar irradiance. This is the error correction value for relative humidity. This represents the error correction value for temperature. Through the above preprocessing and correction, high-quality, high-resolution corrected NWP data can be obtained, providing reliable meteorological input for the two-stage deep learning prediction model.

[0061] In this embodiment, the deep learning model makes predictions by constructing a mapping relationship between future photovoltaic power output and relevant input factors. The overall timeline is defined as the historical time step. With future time steps Union of: Based on a given historical photovoltaic power Measured meteorological data NWP output and model parameters W The goal is to predict photovoltaic power output. In this embodiment, this can be transformed into an optimization problem aimed at determining the conditional sequence for the next 96 time steps:

[0062]

[0063] in, Represents the prediction output matrix of different models, including , and These represent the predicted outputs of the LSTM mapping model, the LSTM time series model, and the CNN model, respectively. It can be represented by a neural network architecture. and These represent the hyperparameters of the LSTM mapping model, LSTM time series model, and CNN model, respectively.

[0064] Considering the differences in the impact of external meteorological factors and internal photovoltaic power generation status on the daily power output time series, this embodiment proposes a Time-Segment Integration Method (TSIM) to optimize the contribution matrix of each model in different time periods. The formula for the time-segmented integration method is:

[0065] in, m This is the total number of time periods. n It represents the number of prediction models. Indicates the first i The predicted power output matrix for each time period.

[0066] To address the overfitting issue of deep learning in daily weather variations, the first stage establishes an LSTM mapping model combining a mapping strategy and NWP technology to predict power trends with strong adaptability. Then, in the second stage, a hybrid deep learning model combining wavelet decomposition, LSTM, and CNN is proposed to extract adjacent and similar features from photovoltaic power on historical adjacent dates and under the same weather conditions, respectively.

[0067] Considering the varying impact of the correlation between NWP output and historical power at different times of the following day, we propose a time-segmented ensemble method. Specifically, we propose a hyperparameter optimization method based on grid search and Bayesian optimization to improve learning efficiency. The prediction framework structure is as follows: Figure 2 As shown.

[0068] Due to the uncertainty of historical NWP information, directly establishing a mapping model between NWP output and historical photovoltaic power output will inevitably lead to severe overfitting. Therefore, a model based on historical measured meteorological data should be used instead. (Solar irradiance) relative humidity and temperature and historical power data To develop and train mapping models, and to use the next day's weather information As input to the model to obtain a preliminary predicted power output Performance is better under clear weather conditions than under other weather conditions, but when applied to other weather conditions, it can only be used as a reference for determining the overall trend.

[0069] In one possible embodiment, in step S2, a mapping model based on a long short-term memory network is established and trained, using historical measured meteorological data and corresponding historical photovoltaic power data as the training set. During the process of learning the mapping relationship between meteorological conditions and power generation, the training and prediction of the mapping relationship satisfy the following formula:

[0070]

[0071] In the formula, This refers to the mapping relationship during the model training phase. This represents the mapping relationship during the model testing phase. For the hyperparameters of the mapping model, , , These are historically measured data for solar irradiance, relative humidity, and temperature. Historical photovoltaic power data, , , These are the predicted solar irradiance, relative humidity, and temperature data for the forecast day. This is the overall trend prediction result.

[0072] In step S2, the first path uses a wavelet decomposition combined with a long short-term memory network model to decompose the historical photovoltaic power sequence and extract the temporal features of the nearest day, outputting the first path fluctuation prediction result. During this process, the wavelet decomposition combined with the long short-term memory network model calculates the following formula:

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] In the formula, , , , , , , This is the weight matrix. , , , For bias vectors, Forget gate activation vector, It is the Sigmoid activation function. Input to Long Short-Term Memory (LSTM) network units, The input gate activation vector, To update the gate output value, It is the ReLU activation function. For scalar multiplication, For the unit state, The output gate activation vector, This is the output of the Long Short-Term Memory (LSTM) network unit.

[0079] In this embodiment, a rough trend of predicted daily photovoltaic power can be obtained using the developed LSTM mapping model. Based on this, a historical feature extraction model is established to make detailed predictions of photovoltaic power fluctuations. The historical feature extraction part includes two prediction models. The first is a WD-LSTM time series model for extracting adjacent features, and the second is a CNN model for extracting similar features.

[0080] To compensate for the inability of the LSTM mapping model to capture adjacent characteristics of historical photovoltaic power and to overcome the strong fluctuations in power output at 15-minute intervals, this embodiment develops a WD-LSTM time series model to analyze the transition characteristics of adjacent dates. Unlike the optimized mapping model based on meteorological information, the WD-LSTM time series model is used to study and analyze historical power trends (especially daily rising and falling periods). WD is used to decompose the original photovoltaic power sequence into main low-frequency and high-frequency sequences.

[0081] Specifically, the historical photovoltaic power sequence is decomposed into subsequences of different frequencies using wavelet decomposition, and the wavelet decomposition satisfies the following formula:

[0082]

[0083] In the formula, It is a low-frequency sequence. , It is a high-frequency sequence. As a scale factor, The translation factor is... This is a historical photovoltaic power sequence. For wavelet basis functions, yes The conjugate function of the decomposed sequence is used; the long short-term memory network model is used to predict the decomposed sequence respectively, and the predicted values ​​are summed to obtain the fine fluctuation prediction results corresponding to the time series characteristics of the nearby days.

[0084] In step S2, the first path uses a wavelet decomposition combined with a long short-term memory network model to decompose the historical photovoltaic power sequence and extract the temporal features of the nearest day. During the output of the first path fluctuation prediction result, the first path fluctuation prediction result corresponding to the temporal features of the nearest day satisfies the following formula:

[0085] In the formula, The prediction results are for the time series features of nearby days. For the time series model of Long Short-Term Memory Network, For model hyperparameters, This is the decomposed historical power subsequence.

[0086] In this embodiment, in addition to mining adjacent features, a CNN model is also used to mine similar features of historical power. Photovoltaic power generation is significantly affected by the uncertainty of weather conditions. Therefore, this study considers the similarity of power trends under different weather conditions. First, the historical data of photovoltaic power generation is clustered into four weather conditions (sunny, cloudy, rainy, and overcast), which can be described as... Furthermore, according to the predicted date Based on the predicted weather conditions, data clusters corresponding to similar historical weather conditions are selected and input into a CNN model to obtain photovoltaic power generation prediction data with similar daily characteristics.

[0087] In this process, historical photovoltaic power data is clustered according to weather type. Based on the weather type of the predicted day, historical data of the corresponding category is selected and input into the convolutional neural network model for training and prediction. This yields the second path fluctuation prediction results corresponding to the weather pattern characteristics of similar days, satisfying the formula:

[0088] In the formula, To predict the forecast results corresponding to the daily weather type, It is a convolutional neural network model. This data represents historical power levels for the same weather type. These are the hyperparameters of the model for the corresponding weather type. To predict the daily weather type, Historical weather types.

[0089] Typically, an internal CNN has three main components: pooling layers, activation layers, and convolutional layers. The information extraction process of the convolutional neural network follows the formula:

[0090] In the formula, For the first Layer Each convolutional kernel is located at... The output, For the first Layer The weights of each convolutional kernel, For the first Layer position The input area, For the first Layer The bias of each convolution kernel This is the activation function.

[0091] In this embodiment, in step S3, a time-segmented integration method is adopted to divide the prediction interval according to the rising, fluctuating, and falling periods of the photovoltaic power generation daily power curve. Dynamic weights are assigned to the overall trend prediction result, the first path fluctuation prediction result, and the second path fluctuation prediction result, and during the weighted fusion process, the time-segmented integration method satisfies the following formula:

[0092]

[0093]

[0094] In the formula, For the first The final predicted value for each time period, , , These are the overall trend model, the nearby day time series feature model, and the similar day weather pattern feature model, respectively, at the [number]th [year]. Weighting coefficients for each time period , , The corresponding models are in the 1st month. Predicted values ​​for a given time period The final overall prediction result is presented. Based on the characteristics of daily photovoltaic power generation, the prediction day is divided into an upward period, a fluctuating period, and a downward period. A general NWP mapping trend provides a trend reference for the entire day. Furthermore, similar features account for a larger proportion in the fluctuating prediction section, while adjacent features account for a larger proportion in the upward and downward prediction sections. Based on the prediction results of each sub-model and the actual photovoltaic power output over the week, a genetic algorithm is used to determine the proportions to minimize the prediction error within the week.

[0095] In this embodiment, in step S4, based on an algorithm combining grid search and Bayesian optimization, the hyperparameters of the mapping model based on a long short-term memory network, the wavelet decomposition combined with a long short-term memory network model, and the convolutional neural network model are automatically optimized. The robustness of the models is evaluated through K-fold cross-validation, and the final day-ahead photovoltaic power prediction result is output. The optimal value of the hyperparameters obtained through Bayesian optimization satisfies the following formula:

[0096] In the formula, The optimal value for the hyperparameter is... For the hyperparameters to be optimized, For the search space of hyperparameters, Let be the objective function. K-fold cross-validation is a method for evaluating the robustness of predictive deep learning models. In this method, the evaluation dataset is repeatedly split into... k A mutually exclusive subset is defined as a set of approximately equal size. The same hyperparameters are used to evaluate the deep learning performance. Finally, the proposed model with optimal parameters is determined by comparing the cross-validation results.

[0097] Specifically, deep learning-based model development and hyperparameter selection is a highly iterative and "black box" process; only the model's input and output can be observed, but gradient information about the model training process cannot be obtained. To assist in predictive model development, this embodiment proposes a systematic and principled model optimization method. Specifically, the optimization process is demonstrated based on grid search and Bayesian optimization to seek the optimal combination of multidimensional parameters of LSTM and CNN models. To verify and evaluate the robustness of the optimized predictive model, K-fold cross-validation is employed.

[0098] To optimize model performance, this invention employs a strategy combining grid search and Bayesian optimization for automatic hyperparameter optimization. First, grid search determines discrete parameters such as optimizer type. Then, Bayesian optimization efficiently searches for optimal values ​​for continuous parameters such as learning rate and training period. K-fold cross-validation is used to evaluate model robustness. Considering the impact of changing weather conditions on power trends at different times, and the fact that short periods of power trend increases and decreases are less affected by weather, while intermediate periods are more severely affected, a time-segmented ensemble method is used to find the optimal proportion of outputs from each proposed model.

[0099] Through field tests on photovoltaic power plants of different sizes, 10 MW and 100 kW, the method proposed in this invention demonstrates excellent precision forecasting capability and strong adaptability in 15-minute high-resolution day-ahead forecasting, and is not limited by the size of the power plant or the weather type. Specifically, in the 10 MW power plant test, the root mean square error (RMSE) of the daily forecast under four typical weather conditions—sunny, cloudy, overcast, and rainy—was 0.18 MW, 0.62 MW, 0.50 MW, and 0.23 MW, respectively, and the mean absolute error (MAE) was 0.10 MW, 0.30 MW, 0.24 MW, and 0.12 MW, respectively. In the 100 kW power plant, the RMSE under the corresponding weather conditions was 1.34 kW, 3.35 kW, 3.52 kW, and 1.72 kW, respectively, and the MAE was 0.86 kW, 1.90 kW, 1.87 kW, and 0.93 kW, respectively. During a 30-day continuous operation test, the prediction accuracy of this method consistently exceeded the 85% baseline required by the power grid dispatching system. The monthly average prediction accuracy for 10 MW and 100 kW power plants reached 93.75% and 93.12%, respectively. These results fully demonstrate that this method possesses high accuracy, strong adaptability, and excellent engineering practicality under various photovoltaic power plant sizes and weather scenarios, effectively supporting the refined power prediction requirements of the power grid's day-ahead dispatching.

[0100] It should be noted that the method of this embodiment can also be applied to distributed scenarios, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method in this embodiment, and the multiple devices will interact with each other to complete the photovoltaic power prediction and analysis method based on two-stage hybrid deep learning.

[0101] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0102] See Figure 3 This invention also provides a photovoltaic power day-ahead prediction device based on two-stage hybrid deep learning, and a photovoltaic power prediction and analysis method based on two-stage hybrid deep learning using the above embodiments or any possible implementation thereof, including: The data preprocessing module 100 is used to perform preprocessing and error correction on the acquired raw numerical weather forecast data to obtain corrected numerical weather forecast data. The two-stage prediction model construction module 200 is used to construct a two-stage hybrid deep learning prediction model framework, which includes an overall trend prediction stage of numerical weather forecast mapping and a fluctuation prediction stage of historical feature extraction. The first-stage processing module 300 is used in the overall trend prediction stage of the numerical weather forecast mapping to establish and train a mapping model based on a long short-term memory network. It uses historical measured meteorological data and corresponding historical photovoltaic power data as training sets to learn the mapping relationship between meteorological conditions and power generation. It inputs the corrected numerical weather forecast data and outputs the overall trend prediction result of photovoltaic power. The second-stage processing module 400 is used in the fluctuation prediction stage of historical feature extraction. It performs dual-path historical feature extraction in parallel. The first path uses a wavelet decomposition combined with a long short-term memory network model to decompose the historical photovoltaic power sequence and extract the temporal features of the adjacent days, and outputs the first-path fluctuation prediction result. The second path uses a convolutional neural network model to extract similar day weather pattern features based on historical data of weather type clustering, and outputs the second-path fluctuation prediction result. The weighted fusion module 500 is used to divide the prediction interval according to the rising, fluctuating and falling periods of the photovoltaic power generation daily power curve using a time segmented integration method, and to assign dynamic weights to the overall trend prediction result, the first path fluctuation prediction result and the second path fluctuation prediction result and then weight and fuse them. The hyperparameter optimization and result output module 600 is used to automatically optimize the hyperparameters of the mapping model based on the long short-term memory network, the wavelet decomposition combined with the long short-term memory network model, and the convolutional neural network model based on an algorithm combining grid search and Bayesian optimization. The module evaluates the robustness of the model through K-fold cross-validation and outputs the final day-ahead photovoltaic power prediction result.

[0103] The apparatus of the above embodiments is used to implement the photovoltaic power prediction and analysis method based on two-stage hybrid deep learning in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0104] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the photovoltaic power prediction and analysis method based on two-stage hybrid deep learning as described in any of the above embodiments.

[0105] Figure 4 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 710, a memory 720, an input / output interface 730, a communication interface 740, and a bus 750. The processor 710, memory 720, input / output interface 730, and communication interface 740 are interconnected internally via the bus 750.

[0106] The processor 710 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0107] The memory 720 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 720 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 720 and is called and executed by the processor 710.

[0108] The input / output interface 730 is used to connect input / output modules to enable information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0109] The communication interface 740 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0110] Bus 750 includes a pathway for transmitting information between various components of the device, such as processor 710, memory 720, input / output interface 730, and communication interface 740.

[0111] It should be noted that although the above-described device only shows the processor 710, memory 720, input / output interface 730, communication interface 740, and bus 750, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0112] The electronic devices described above are used to implement the photovoltaic power prediction and analysis method based on two-stage hybrid deep learning in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0113] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, the present invention also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the photovoltaic power prediction and analysis method based on two-stage hybrid deep learning as described in any of the above embodiments.

[0114] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0115] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the photovoltaic power prediction and analysis method based on two-stage hybrid deep learning as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0116] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the embodiments of the invention as described above, which are not provided in detail for the sake of brevity.

[0117] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of the invention, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of the invention, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of the invention will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of the invention, it will be apparent to those skilled in the art that the embodiments of the invention may be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0118] Although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., DRAM) may use the embodiments discussed.

[0119] The embodiments of this invention are intended to cover all such substitutions, modifications, and variations falling within the scope of the claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this invention should be included within the scope of protection of this invention.

Claims

1. A photovoltaic power prediction and analysis method based on two-stage hybrid deep learning, characterized in that, Includes the following steps: Preprocessing and error correction are performed on the acquired raw numerical weather forecast data to obtain corrected numerical weather forecast data; A two-stage hybrid deep learning prediction model framework is constructed, which includes an overall trend prediction stage based on numerical weather prediction mapping and a fluctuation prediction stage based on historical feature extraction. In the overall trend prediction stage of the numerical weather forecast mapping, a mapping model based on a long short-term memory network is established and trained. Using historical measured meteorological data and corresponding historical photovoltaic power data as training sets, the mapping relationship between meteorological conditions and power generation is learned. The corrected numerical weather forecast data is input, and the overall trend prediction result of photovoltaic power is output. In the fluctuation prediction stage of historical feature extraction, dual-path historical feature extraction is performed in parallel. The first path uses a wavelet decomposition combined with a long short-term memory network model to decompose the historical photovoltaic power sequence and extract the temporal features of adjacent days, outputting the first path fluctuation prediction result. The second path uses a convolutional neural network model to extract similar day weather pattern features based on historical data of weather type clustering, outputting the second path fluctuation prediction result. A time-segmented integration method is adopted, and the prediction interval is divided according to the rising, fluctuating and falling periods of the photovoltaic power generation daily power curve. Dynamic weights are assigned to the overall trend prediction result, the first path fluctuation prediction result and the second path fluctuation prediction result and then weighted and integrated. An algorithm combining grid search and Bayesian optimization is used to automatically optimize the hyperparameters of the mapping model based on long short-term memory network, the wavelet decomposition combined with long short-term memory network model, and the convolutional neural network model. The robustness of the model is evaluated by K-fold cross-validation, and the final day-ahead photovoltaic power prediction result is output.

2. The two-stage hybrid deep learning based photovoltaic power prediction analysis method according to claim 1, characterized in that, The process of preprocessing and error correction on the acquired raw numerical weather prediction data to obtain corrected numerical weather prediction data specifically includes: Interpolation methods are used to interpolate the raw numerical weather forecast data to match the time resolution of the photovoltaic power data; Based on historical measured meteorological datasets and historical numerical weather prediction datasets, a time-by-time error correction method is used to reduce prediction errors. The time-by-time error correction process satisfies the following formula: ； In the formula, For corrected numerical weather forecast data, The original numerical weather forecast data to be corrected. This is the error correction value. For the first Heavenly The error value between numerical weather forecast data and measured data at a given time point For the first Error weighting for each day It is a set of days for historical data. For the first time in history Heavenly Numerical weather forecast data at specific points in time, For the first time in history Heavenly Measured meteorological data at specific time points; said error correction value is a set of correction values for solar irradiance, relative humidity, and air temperature, satisfying the formula: ； In the formula, is an error correction value for solar irradiance, is an error correction value for relative humidity, is an error correction value for air temperature. 3.The two-stage hybrid deep learning based photovoltaic power forecasting analysis method according to claim 1, wherein, A mapping model based on a long short-term memory network is established and trained. Using historical measured meteorological data and corresponding historical photovoltaic power data as the training set, the mapping relationship between meteorological conditions and power generation is learned. During this process, the training and prediction of this mapping relationship satisfy the following formula: ； ； In the formula, This refers to the mapping relationship during the model training phase. This represents the mapping relationship during the model testing phase. For the hyperparameters of the mapping model, , , These are historically measured data for solar irradiance, relative humidity, and temperature. Historical photovoltaic power data, , , These are the predicted solar irradiance, relative humidity, and temperature data for the forecast day. This is the overall trend prediction result.

4. The two-stage hybrid deep learning based photovoltaic power prediction analysis method according to claim 3, characterized in that, The first path utilizes a wavelet decomposition combined with a long short-term memory network model to decompose historical photovoltaic power sequences and extract time-series features of nearby days. During the output of the first path fluctuation prediction results, the wavelet decomposition combined with the long short-term memory network model calculates according to the following formula: ； ； ； ； ； ； In the formula, , , , , , , This is the weight matrix. , , , For bias vectors, Forget gate activation vector, It is the Sigmoid activation function. Input to Long Short-Term Memory (LSTM) network units, The input gate activation vector, To update the gate output value, It is the ReLU activation function. For scalar multiplication, For the unit state, The output gate activation vector, For output of Long Short-Term Memory (LSTM) network units; Extracting nearby day time series features after decomposing historical photovoltaic power sequences specifically includes: The historical photovoltaic power sequence is decomposed into subsequences of different frequencies using wavelet decomposition, which satisfies the following formula: ； ； In the formula, It is a low-frequency sequence. , It is a high-frequency sequence. As a scale factor, The translation factor is... This is a historical photovoltaic power sequence. For wavelet basis functions, yes The conjugate function of the decomposed sequence is used; the long short-term memory network model is used to predict the decomposed sequence respectively, and the predicted values ​​are summed to obtain the fine fluctuation prediction results corresponding to the time series characteristics of the nearby days.

5. The two-stage hybrid deep learning based photovoltaic power prediction analysis method according to claim 4, characterized in that, The first path utilizes a model combining wavelet decomposition and long short-term memory networks to decompose historical photovoltaic power sequences and extract nearby daily time-series features. During the output of the first path fluctuation prediction results, the first path fluctuation prediction results corresponding to the nearby daily time-series features satisfy the following formula: ； In the formula, The prediction results are for the time series features of nearby days. For the time series model of Long Short-Term Memory Network, For model hyperparameters, This is the decomposed historical power subsequence. 6.The two-stage hybrid deep learning based photovoltaic power forecasting analysis method according to claim 1, wherein, The second path utilizes a convolutional neural network model to extract similar day weather pattern features based on historical weather type clustering data, and outputs the second path fluctuation prediction results: Historical photovoltaic power data is clustered and categorized according to weather type. Based on the weather type of the predicted day, historical data of the corresponding category are selected and input into a convolutional neural network model for training and prediction. This yields the second path fluctuation prediction results corresponding to the weather pattern characteristics of similar days, satisfying the formula: ； In the formula, To predict the forecast results corresponding to the daily weather type, It is a convolutional neural network model. This data represents historical power levels for the same weather type. These are the hyperparameters of the model for the corresponding weather type. To predict the daily weather type, Historical weather types.

7. The two-stage hybrid deep learning based photovoltaic power prediction analysis method according to claim 6, characterized in that, The information extraction process of the convolutional neural network satisfies the following formula: ； In the formula, For the first Layer Each convolutional kernel is located at... The output, For the first Layer The weights of each convolutional kernel, For the first Layer position The input area, For the first Layer The bias of each convolution kernel This is the activation function. 8.The two-stage hybrid deep learning based photovoltaic power forecasting analysis method according to claim 1, wherein, The time-segmented integration method divides the prediction interval according to the rising, fluctuating, and falling periods of the photovoltaic power generation daily power curve. Dynamic weights are assigned to the overall trend prediction result, the first path fluctuation prediction result, and the second path fluctuation prediction result during the weighted fusion process. The time-segmented integration method satisfies the following formula: ； ； ； In the formula, For the first The final predicted value for each time period, , , These are the overall trend model, the nearby day time series feature model, and the similar day weather pattern feature model, respectively, at the [number]th [year]. Weighting coefficients for each time period , , The corresponding models are in the 1st month. Predicted values ​​for a given time period This represents the final overall prediction result.

9. The photovoltaic power prediction and analysis method based on two-stage hybrid deep learning according to claim 1, characterized in that, An algorithm combining grid search and Bayesian optimization is used to automatically optimize the hyperparameters of the mapping model based on a long short-term memory network, the wavelet decomposition combined with a long short-term memory network model, and the convolutional neural network model. The robustness of the models is evaluated through K-fold cross-validation, and the final day-ahead photovoltaic power prediction result is output. The optimal value of the hyperparameters obtained through Bayesian optimization satisfies the following formula: ； In the formula, The optimal value for the hyperparameter is... For the hyperparameters to be optimized, For the search space of hyperparameters, The objective function is denoted as .

10. A photovoltaic power day-ahead forecasting device based on two-stage hybrid deep learning, adopting the photovoltaic power forecasting analysis method based on two-stage hybrid deep learning according to any one of claims 1 to 9, characterized in that, include: The data preprocessing module is used to perform preprocessing and error correction on the acquired raw numerical weather forecast data to obtain corrected numerical weather forecast data. A two-stage prediction model construction module is used to construct a two-stage hybrid deep learning prediction model framework, which includes an overall trend prediction stage based on numerical weather forecast mapping and a fluctuation prediction stage based on historical feature extraction. The first-stage processing module is used in the overall trend prediction stage of the numerical weather forecast mapping to establish and train a mapping model based on a long short-term memory network. It uses historical measured meteorological data and corresponding historical photovoltaic power data as training sets to learn the mapping relationship between meteorological conditions and power generation. It inputs the corrected numerical weather forecast data and outputs the overall trend prediction result of photovoltaic power. The second-stage processing module, used in the fluctuation prediction stage of historical feature extraction, performs parallel dual-path historical feature extraction. The first path uses a wavelet decomposition combined with a long short-term memory network model to decompose the historical photovoltaic power sequence and extract the temporal features of adjacent days, outputting the first path fluctuation prediction result. The second path uses a convolutional neural network model to extract similar day weather pattern features based on historical data of weather type clustering, outputting the second path fluctuation prediction result. The weighted fusion module is used to divide the prediction interval according to the rising, fluctuating and falling periods of the photovoltaic power generation daily power curve using a time segmented integration method, and to assign dynamic weights to the overall trend prediction result, the first path fluctuation prediction result and the second path fluctuation prediction result and then weighted fusion them. The hyperparameter optimization and result output module is used to automatically optimize the hyperparameters of the mapping model based on long short-term memory network, the wavelet decomposition combined with long short-term memory network model, and the convolutional neural network model based on an algorithm combining grid search and Bayesian optimization. The module evaluates the robustness of the model through K-fold cross-validation and outputs the final day-ahead photovoltaic power prediction result.

Images