A method and system for predicting ultra-short-term photovoltaic power output based on parallel data processing

Abstract

This invention discloses an ultra-short-term photovoltaic (PV) power output prediction method and system based on parallel data processing. The prediction method includes: acquiring meteorological forecast data for a first preset time period; executing the following three steps in parallel through three independent processes: statistically analyzing PV power output prediction data based on the meteorological forecast data; obtaining mechanistic PV power output prediction data based on the PV panel output characteristics and the meteorological forecast data; merging the statistical PV power output prediction data, the mechanistic PV power output prediction data, and historical meteorological data and historical PV power output data for a second preset time period into an input dataset, training a convolutional neural network, and generating the final predicted PV power output data. This application meets the real-time requirements of ultra-short-term prediction in terms of efficiency, reduces errors in terms of accuracy through multi-model fusion and deep data utilization, and adapts to the actual operational needs of PV power plants in terms of practicality, providing a reliable technical solution for ultra-short-term PV power output prediction.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power output prediction technology, and in particular to an ultra-short-term photovoltaic power output prediction method and system based on parallel data processing. Background Technology

[0002] With the global energy structure shifting towards clean energy, the photovoltaic (PV) industry, as a crucial component of the renewable energy sector, has experienced rapid development, with its installed capacity and power generation continuously increasing. As an energy output form dependent on natural conditions, PV power output is significantly affected by meteorological factors such as solar radiation, temperature, and cloud movement, exhibiting strong intermittency and volatility. This characteristic poses significant challenges to the safe and stable operation of power systems, the optimal scheduling of power resources, and the improvement of energy utilization efficiency.

[0003] Accurate photovoltaic (PV) output forecasting is a key technological support for solving the aforementioned problems. Among these, ultra-short-term PV output forecasting can provide immediate decision-making basis for real-time grid dispatching, electricity market trading, and coordinated control of energy storage systems, playing an irreplaceable role in maintaining power supply and demand balance and reducing curtailment rates. However, current ultra-short-term PV output forecasting technology still faces the following prominent problems:

[0004] 1. Traditional forecasting methods struggle to balance speed and accuracy.

[0005] When faced with drastic fluctuations in photovoltaic output caused by sudden weather changes (such as sudden cloud cover or short-term strong radiation), traditional forecasting methods either suffer from insufficient accuracy due to low model complexity and reliance on historical data trends, failing to capture instantaneous change patterns; or they prolong computation time due to the use of complex models (such as traditional machine learning algorithms), making it difficult to meet the core requirement of "real-time" for ultra-short-term forecasting (which typically requires forecasting and outputting results within minutes), and thus failing to provide rapid response support for grid dispatch.

[0006] 2. The correlations among multi-dimensional data have not been fully explored.

[0007] Traditional forecasting methods do not conduct a synergistic analysis of multiple factors such as meteorological conditions, equipment characteristics, and station operating status, resulting in significant deviations between the forecast results and the actual situation. Consequently, they fail to meet the stringent requirements of power systems for ultra-short-term photovoltaic power output forecasting.

[0008] In summary, the shortcomings of existing ultra-short-term photovoltaic power output prediction methods in terms of speed, accuracy, and multi-dimensional data utilization have become bottlenecks restricting the efficient consumption of photovoltaic power and the stable operation of the power system. Summary of the Invention

[0009] In view of the technical problems mentioned in the background, the purpose of this invention is to provide an ultra-short-term photovoltaic power output prediction method and system based on parallel data processing.

[0010] To achieve the objectives of this invention, the technical solution provided by this invention is as follows:

[0011] First rise

[0012] This application provides a method for predicting ultra-short-term photovoltaic power output based on data parallel processing, including the following steps:

[0013] Step S1: Obtain meteorological forecast data for the first preset time period;

[0014] Step S2: Execute the following three steps in parallel using three independent processes:

[0015] Step S2-1: Based on the meteorological forecast data, extract historical similar weather data from historical meteorological data, and obtain corresponding historical photovoltaic power output data based on the historical similar weather data; use the historical similar weather data and historical photovoltaic power output data as a training set to train the XGBoost model; input the meteorological forecast data into the trained XGBoost model to obtain statistical photovoltaic power output forecast data;

[0016] Step S2-2: Based on the output characteristics of photovoltaic panels and meteorological forecast data, obtain the mechanism photovoltaic output prediction data;

[0017] Step S2-3: Combine the statistical photovoltaic power output prediction data, the mechanistic photovoltaic power output prediction data, and the historical meteorological data and corresponding historical photovoltaic power output data within the second preset time period into an input dataset, train a convolutional neural network, and generate the final predicted photovoltaic power output data.

[0018] Second aspect

[0019] Corresponding to the above method, this application provides an ultra-short-term photovoltaic power output prediction system based on parallel data processing, including the following units: a meteorological forecast data acquisition unit and a parallel processing unit;

[0020] The meteorological forecast data acquisition unit is used to acquire meteorological forecast data within a first preset time period;

[0021] The parallel processing unit is used to execute the following three steps in parallel through three independent processes:

[0022] Step S2-1: Based on the meteorological forecast data, extract historical similar weather data from historical meteorological data, and obtain corresponding historical photovoltaic power output data based on the historical similar weather data; use the historical similar weather data and historical photovoltaic power output data as a training set to train the XGBoost model; input the meteorological forecast data into the trained XGBoost model to obtain statistical photovoltaic power output forecast data;

[0023] Step S2-2: Based on the output characteristics of photovoltaic panels and meteorological forecast data, obtain the mechanism photovoltaic output prediction data;

[0024] Step S2-3: Combine the statistical photovoltaic power output prediction data, the mechanistic photovoltaic power output prediction data, and the historical meteorological data and corresponding historical photovoltaic power output data within the second preset time period into an input dataset, train a convolutional neural network, and generate the final predicted photovoltaic power output data.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] I. Improve the efficiency of ultra-short-term forecasting to meet real-time requirements.

[0027] The prediction method provided in this application adopts a data-parallel processing architecture, which executes statistical prediction, mechanistic prediction, and fusion prediction steps in parallel through three independent processes, avoiding the problem of "preceding steps blocking subsequent steps" in traditional serial computing. The parallel processing mode significantly shortens the overall prediction time, ensuring that prediction results can be generated in a short time, providing timely support for subsequent power dispatching decisions.

[0028] II. Integrating multi-dimensional prediction logic to improve prediction accuracy

[0029] The use of multiple complementary models reduces single errors: Statistical prediction (XGBoost) is trained on historical similar weather data and is good at capturing data correlation patterns; Mechanistic prediction is based on the physical characteristics of photovoltaic panels and meteorological parameters, which is consistent with the physical nature of photovoltaic power output; Convolutional Neural Network (CNN) further integrates the two types of prediction results and historical data, and can mine cross-dimensional features (such as the correlation between the temporal changes of meteorological parameters and power output fluctuations). The combination of the three avoids the limitations of a single model.

[0030] Comprehensive meteorological data supports accurate input: Meteorological forecast data covers 10 core parameters, including weather type and irradiance, and includes detailed information such as sunrise and sunset times and second-level timestamps, providing ample input for various models. For example, parameters such as irradiance and photovoltaic irradiance angle directly affect the calculation accuracy of the mechanistic model, while data such as visibility and relative humidity from similar historical weather conditions improve the adaptability of the XGBoost model to complex weather conditions.

[0031] III. Enhance the depth of data utilization and improve the robustness of predictions.

[0032] Multi-source data fusion to uncover latent features: The input dataset of the convolutional neural network integrates statistical prediction results, mechanistic prediction results, historical meteorological data, and historical power output data, constructing a time-series feature sequence through a 200×10 feature matrix. This fusion not only utilizes the direct correlation between "current weather and current power output" but also captures the latent pattern of "power output trend on similar historical days - current prediction deviation," making the model's predictions of sudden weather changes (such as short-term cloud cover) more stable.

[0033] Similar day data enhances generalization ability: Training with data from the same period of historically similar weather days reduces interference from dissimilar weather data. For example, the photovoltaic power output patterns differ significantly between sunny and cloudy days. Filtering with similar day data makes the training samples for the XGBoost model more targeted, improving the model's predictive adaptability under specific weather types.

[0034] IV. Adaptable to ultra-short-term scenarios, supporting efficient operation of photovoltaic power plants

[0035] Ultra-short-term (within 4 hours) photovoltaic power output is greatly affected by real-time weather changes (such as sudden showers and sudden changes in wind speed). Traditional methods often fail to meet the requirements due to their slow response and large errors. This method adapts to the scenario requirements through the following design: parallel processing ensures rapid generation of prediction results, adapting to the need for "short-term adjustment"; the mechanism model is based on the physical characteristics of photovoltaic panels (such as conversion coefficient and irradiation angle), which responds more directly to sudden weather changes; the temporal feature capture capability of CNN can identify the correlation between "small fluctuations and power output changes" in parameters such as irradiance and temperature, and predict the trend of power output fluctuations in advance.

[0036] In summary, the technical solution provided in this application meets the real-time requirements of ultra-short-term forecasting in terms of efficiency, reduces errors through multi-model fusion and in-depth data utilization in terms of accuracy, and adapts to the actual operation needs of photovoltaic power plants in terms of practicality, thus providing a reliable technical solution for ultra-short-term photovoltaic power output forecasting. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0039] like Figure 1 As shown, this embodiment provides an ultra-short-term photovoltaic power output prediction method based on data parallel processing, including the following steps:

[0040] Step S1: Obtain meteorological forecast data for the first preset time period;

[0041] In step S1, the first preset time period is 4 hours from the predicted time point.

[0042] In step S1, the meteorological forecast data includes weather type, wind speed, wind direction, temperature, relative humidity, visibility, air pressure, irradiance, and sunrise and sunset times. The meteorological forecast data is generated by the meteorological forecast system of the photovoltaic power station.

[0043] All the above data were collected every 10 seconds, and after data cleaning and other preprocessing steps, the data was stored in a local database.

[0044] In addition, non-numerical meteorological forecast data are coded, and the weather types are mainly divided into 7 types: sunny (7), cloudy (6), overcast (5), fog (4), light rain / light snow (3), moderate rain / moderate snow (2), heavy rain / heavy snow (1); the wind direction is mainly divided into 8 types such as northerly wind (0) and northeasterly wind (1); the sunrise and sunset times are divided into: sunrise (1) and sunset (0).

[0045] Step S2: Execute the following three steps in parallel using three independent processes:

[0046] Step S2-1: Based on the meteorological forecast data, extract historical similar weather data from historical meteorological data, and obtain corresponding historical photovoltaic power output data based on the historical similar weather data; use the historical similar weather data and historical photovoltaic power output data as a training set to train the XGBoost model; input the meteorological forecast data into the trained XGBoost model to obtain statistical photovoltaic power output forecast data;

[0047] Specifically, features are extracted from meteorological forecast data to generate weather feature sequences; based on these weather feature sequences, historically similar weather data are extracted from historical meteorological data in the database to generate a highly correlated historical dataset. The historical meteorological data refers to meteorological data prior to the forecast time point and is stored in the database of the photovoltaic power station.

[0048] It should be noted that principal component analysis is used to extract characteristic values ​​and eigenvectors from meteorological forecast data. The specific process is as follows:

[0049] First, anomaly processing is performed on the meteorological forecast data to remove potential systematic trends in meteorological elements. A covariance matrix is ​​then generated based on the processed data, and eigenvalues ​​and eigenvectors are extracted. The eigenvalues ​​reflect the importance of the corresponding eigenvectors within the meteorological forecast data.

[0050] Secondly, using meteorological forecast data as the target meteorological data, and to accurately select historical meteorological data samples with strong correlation to the target meteorological data, the grey relational analysis method from grey system theory is introduced to systematically calculate the correlation degree between the weather feature sequences of the target meteorological data and the weather feature sequences of each historical meteorological data. The weather feature sequence of the target meteorological data is defined as the reference sequence, and the weather feature sequences of all historical meteorological data are used as comparison sequences. Mean or range processing is used to eliminate interference caused by differences in the dimensions and orders of magnitude of different meteorological parameters (such as irradiance, temperature, wind speed, etc.). Then, the correlation coefficient matrix between the reference sequence and each comparison sequence is solved one by one according to the grey relational degree calculation formula. Finally, a weighted average is used to obtain the comprehensive grey relational degree value, which characterizes the overall correlation degree between the two sets of sequences.

[0051] All historical meteorological data were sorted from highest to lowest correlation value, and the top 10% of highly correlated data were selected to form a highly correlated historical dataset.

[0052] Finally, the aforementioned highly correlated historical meteorological data and corresponding historical photovoltaic power output data were merged into a training set, while the target meteorological data was used as a test set. During the model training phase, the XGBoost model was used to learn and fit the training set data. To achieve global optimization of model parameters, the parameter grid of the XGBoost model needed to be pre-set: the learning rate was set to [0.1, 0.01, 0.001], the maximum tree depth was set to [3, 4, 5], and the subsampling rate was set to [0.8, 0.9, 1.0]. Auxiliary parameters such as the minimum leaf node sample weight and gamma value were also set to form a complete parameter optimization grid. For this parameter grid, a ten-fold cross-validation strategy was used for parameter optimization. The mean of ten validation results was used as the performance evaluation index for this set of parameters. Finally, the optimal parameter combination that minimizes the validation error was selected, ensuring that the model achieves a globally optimal fit during the training phase.

[0053] After model training is completed, the test results are evaluated: the model's prediction accuracy and generalization ability are comprehensively verified by calculating multiple dimensions such as mean squared error, mean absolute error, root mean square error, and mean absolute percentage error. If the evaluation results do not reach the preset accuracy threshold, a secondary parameter optimization process is initiated: expanding the parameter search range (extending the learning rate to [0.2, 0.1, 0.01]) or increasing the training sample size (supplementing historical data with the second highest correlation), etc., and re-executing the training process until the error requirements are met.

[0054] Once the model performance meets the requirements, the XGBoost model structure and weight parameters with the optimal parameter configuration are saved to local storage, achieving persistent storage and reuse of the model. Finally, the target meteorological data feature sequence of the test set is input into the optimized model, and the corresponding statistical photovoltaic output prediction data is generated through the model's forward computation.

[0055] Step S2-2: Based on the photovoltaic panel output characteristics and meteorological forecast data, obtain the mechanism photovoltaic output prediction data; in step S2-2, the calculation formula for the mechanism photovoltaic output prediction data is as follows:

[0056] ;

[0057] in, For the mechanism of photovoltaic power output prediction data, For irradiation intensity, Sunrise and sunset encoded values, The maximum power tracking point, This is a photovoltaic radiation area. For conversion factors, This refers to the photovoltaic irradiation angle.

[0058] Step S2-3: Combine the statistical photovoltaic power output prediction data, the mechanistic photovoltaic power output prediction data, and the historical meteorological data and corresponding historical photovoltaic power output data within the second preset time period into an input dataset, train a convolutional neural network, and generate the final predicted photovoltaic power output data.

[0059] Steps S2-3 specifically include the following:

[0060] Step S2-3-1: For each prediction time point within the prediction time domain, perform the following operations sequentially to form multiple subsets of data, and then merge these subsets into the input dataset:

[0061] Statistical photovoltaic (PV) output prediction data and mechanistic PV output prediction data from the same time points on 200 historical similar weather days, along with historical meteorological data and corresponding historical PV output data within a second preset time period, were sequentially concatenated. Each time point contained 10 weather feature attributes. After concatenation, a 200×10 dimension original feature matrix was formed, which served as the input data for the convolutional neural network input layer. The 10 weather attributes included a timestamp accurate to the second, coded weather type, wind speed, wind direction coded value, temperature, relative humidity, visibility, air pressure, irradiance, and sunrise / sunset coded values. One output attribute constituted a 1×1 output vector, representing the historical PV output data at the corresponding time point, which served as the target value for the convolutional neural network parameter correction.

[0062] Step S2-3-2: Train a convolutional neural network based on the input dataset to obtain the final prediction result consisting of the prediction values ​​at multiple prediction time points.

[0063] The second preset time period is 4 hours before the predicted time point.

[0064] During the model training phase, the local feature extraction capabilities of convolutional neural networks (CNNs) are utilized, and the convolutional kernel parameters and pooling windows are specifically designed for the input data of photovoltaic power output prediction. The specific structure of the convolutional neural network is as follows:

[0065] Input layer: Data after preprocessing (including outlier removal, feature standardization, and temporal alignment) and filtering and splicing, with a channel size of 200x10.

[0066] Convolutional Layer 1: Kernel size is 3@4x1, stride is 1, and output channel size is 3@197x10. The window size is 4x1, with 3 windows forming one convolutional kernel. The kernel size of 3@4x1 is set to treat 4 days of meteorological data as a local region of interest. The stride is set to 1 to slide with the smallest possible step, thus preserving more local detail. The output channel size is set to 3@91x4 to ensure that the most detail is preserved in the initial processing for subsequent use by the neural network.

[0067] Pooling Layer 1: Max pooling, with a window size of 2x2 and an output channel size of 3@91x4. The pooling operation compresses the values ​​of each 2x2 region into a single value, and uses max pooling to select the maximum value, preserving the most important features of that region, enhancing the model's robustness to input noise, and reducing redundant computation.

[0068] Convolutional Layer 2: Kernel size is 3@8x2, stride is 1, and output channel size is 3@91x4. Compared to Convolutional Layer 1, this layer has a larger kernel size of 3@8x2. The purpose is to broaden the local focus area, ignore some details, avoid overfitting, and achieve deep extraction of intermediate-order correlation features.

[0069] Pooling layer 2: max pooling, window size is 2x2, output channel size is 3@45x2.

[0070] Convolutional Layer 3: Kernel size 3@2x1, stride 1, output channel size 3@44x2. Compared to the first two convolutional layers, the smaller kernel size is used to refine the intermediate-level features extracted earlier, focusing on capturing subtle changes in the feature mapping. Convolutional layers 1, 2, and 3 decrease in the number of channels from 200, 197, and 91 to 44. Through the progressive processing of three consecutive convolutional layers, the transformation from low-level meteorological features (raw irradiance, temperature values) to high-level correlated features (nonlinear mapping law between meteorology and photovoltaic power output) is gradually completed, achieving effective compression of feature dimensions.

[0071] Pooling layer 3: max pooling, window size is 2x2, output channel size is 3@22x1.

[0072] The flatten layer has an output channel size of 66x1. It flattens the input data and transforms the two-dimensional feature mapping into a one-dimensional feature sequence through dimensionality transformation.

[0073] Fully connected layer: The fully connected layer adopts a three-level progressive structure: The first layer has 128 neurons to perform preliminary nonlinear combination of the flat 66-length feature vectors and explore the implicit correlation between cross features; the second layer is reduced to 64 neurons to further refine the core feature patterns and reduce the model complexity; the third layer is compressed to 3 neurons to enhance the strong correlation between the output and photovoltaic power output.

[0074] Output layer: Through linear transformation, the 1-dimensional 3x1 vector features of the fully connected layer are mapped to a 1x1 single value, which is the final predicted maximum photovoltaic output value, thus achieving an accurate mapping from meteorological feature data to a single output prediction value.

[0075] The convolutional neural network model deeply mines the temporal correlation features (such as the coupling relationship of power output at similar times) in the input dataset and the complementary features of multi-source prediction data. After multi-layer nonlinear mapping and feature recombination, the final prediction result of photovoltaic power output is finally generated.

[0076] It should be noted that this application significantly shortens the overall prediction time by decomposing the model computation task into independent processes for parallel execution. The collected data, after standardization and preprocessing, forms a feature data package that can be computed in parallel, and is synchronously distributed to two independent process units: Process 1 (step S2-1) and Process 2 (step S2-2). Process 3 (step S2-3) receives the intermediate prediction results output by Process 1 and Process 2. The three processes are independent in their allocation of computing resources and can be synchronously advanced using the parallel computing capabilities of multi-core processors. The computation results of Process 3 are then concatenated through time-series alignment to generate the final prediction sequence. The operation flow of this parallel computing architecture follows the core logic of "data distribution - parallel computation - result aggregation".

[0077] To verify the efficiency improvement effect of the parallel computing architecture, a comparative experiment was conducted using the above data and the same hardware environment and test dataset. The experimental results show that in the serial computing mode (three models executed sequentially), the total time for a single prediction is 26.73 seconds; however, after enabling process pool parallel computing, because the computational tasks of the three models are executed synchronously on a multi-core processor, the total time is reduced to 10.47 seconds, resulting in a computational efficiency improvement of approximately 60.83%. This result fully verifies the effectiveness of the proposed parallel algorithm in shortening ultra-short-term prediction time, and it is particularly suitable for photovoltaic power plant scheduling scenarios with stringent real-time requirements.

[0078] In addition, corresponding to the above method, this application also provides an ultra-short-term photovoltaic power output prediction system based on parallel data processing, including the following units: a meteorological forecast data acquisition unit and a parallel processing unit;

[0079] The meteorological forecast data acquisition unit is used to acquire meteorological forecast data within a first preset time period;

[0080] The parallel processing unit is used to execute the following three steps in parallel through three independent processes:

[0081] Step S2-1: Based on the meteorological forecast data, extract historical similar weather data from historical meteorological data, and obtain corresponding historical photovoltaic power output data based on the historical similar weather data; use the historical similar weather data and historical photovoltaic power output data as a training set to train the XGBoost model; input the meteorological forecast data into the trained XGBoost model to obtain statistical photovoltaic power output forecast data;

[0082] Step S2-2: Based on the output characteristics of photovoltaic panels and meteorological forecast data, obtain the mechanism photovoltaic output prediction data;

[0083] Step S2-3: Combine the statistical photovoltaic power output prediction data, the mechanistic photovoltaic power output prediction data, and the historical meteorological data and corresponding historical photovoltaic power output data within the second preset time period into an input dataset, train a convolutional neural network, and generate the final predicted photovoltaic power output data.

[0084] Preferably, the first preset time period is 4 hours from the predicted time point.

[0085] Preferably, the meteorological forecast data includes weather type, wind speed, wind direction, temperature, relative humidity, visibility, air pressure, irradiance, and sunrise and sunset times.

[0086] Finally, it should be noted that the above embodiments are merely illustrative and explanatory of the present invention, and are not intended to limit the present invention to the scope of the described embodiments. Furthermore, those skilled in the art will understand that the present invention is not limited to the above embodiments, and many more variations and modifications can be made based on the teachings of the present invention, all of which fall within the scope of protection claimed by the present invention.

Claims

1. A method for ultra-short-term photovoltaic power output prediction based on data parallel processing, characterized in that, The method comprises the following steps: Step S1: obtaining meteorological prediction data in a first preset time period; Step S2: performing the following three steps in parallel through three independent processes: Step S2-1: based on the meteorological prediction data, extracting historical similar weather data from historical meteorological data, and obtaining corresponding historical photovoltaic output data based on the historical similar weather data; training an XGBoost model with the historical similar weather data and historical photovoltaic output data as a training set; inputting the meteorological prediction data into the trained XGBoost model to obtain statistical photovoltaic output prediction data; Step S2-2: obtaining mechanism photovoltaic output prediction data according to photovoltaic panel output characteristics and meteorological prediction data; Step S2-3: merging the statistical photovoltaic output prediction data, the mechanism photovoltaic output prediction data, and historical meteorological data and corresponding historical photovoltaic output data in a second preset time period into an input data set, training a convolutional neural network, and generating final predicted photovoltaic output data; In step S1, the first preset time period is 4 hours from the prediction time point; In step S1, the meteorological prediction data includes weather type, wind speed, wind direction, temperature, relative humidity, visibility, air pressure, irradiance, sunrise and sunset time; In step S2-2, the calculation formula of the mechanism photovoltaic output prediction data is as follows: P PV = rgη max A PV η PV cosθ where P PV is the mechanistic photovoltaic power prediction data, r is the irradiance, g is the sunrise sunset encoding value, η max is the maximum power point, A PV is the photovoltaic radiation area, η PV is the conversion factor, and θ is the photovoltaic incidence angle. Step S2-3, specifically comprising the following: Step S2-3-1: for each prediction time point in the prediction time domain, the following operations are performed in sequence to form multiple sub-data sets, and the multiple sub-data sets are merged into an input data set: selecting a preset number of statistical photovoltaic output prediction data, mechanism photovoltaic output prediction data, and historical meteorological data and corresponding historical photovoltaic output data in a second preset time period of the same time point of the historical similar weather day for sequence splicing, each time point contains 10 weather feature attributes, and after splicing, an original feature matrix with a dimension of 200x10 is formed as input data of the input layer of the convolutional neural network. 10 weather attributes include time stamp accurate to seconds, encoded weather type, wind speed, wind direction encoding value, temperature, relative humidity, visibility, air pressure, irradiance, sunrise and sunset encoding value; 1 output attribute constitutes a 1x1 output vector, representing the historical photovoltaic output data of the corresponding time point, as the result target value of the parameter correction of the convolutional neural network; Step S2-3-2: training the convolutional neural network based on the input data set to obtain a final prediction result composed of prediction values of multiple prediction time points; The preset number is 200, and the second preset time period is 4 hours before the prediction time point.

2. The method according to claim 1, wherein, The convolutional neural network comprises: The input layer has a channel size of 200x10; The convolutional layer 1 has a convolution kernel size of 3@4x1, a step size of 1, and an output channel size of 3@197x10; The pooling layer 1 has a maximum pooling window size of 2x2, and an output channel size of 3@91x4; The convolutional layer 2 has a convolution kernel size of 3@8x2, a step size of 1, and an output channel size of 3@91x4; Pooling layer 2, max pooling, window size 2x2, output channel size 3@45x2; Convolution layer 3, convolution kernel size 3@2x1, step size 1, output channel size 3@44x2; Pooling layer 3, max pooling, window size 2x2, output channel size 3@22x1; Flatten layer, output channel size 66x1, the input data is flattened, and the conversion from two-dimensional feature mapping to one-dimensional feature sequence is realized through dimension conversion; The full connection layer adopts a three-level progressive structure; the first layer is provided with 128 neurons, the 66-length feature vector after flattening is preliminarily combined in a nonlinear manner, and the hidden correlation between cross-features is mined; the second layer is reduced to 64 neurons; and the third layer is compressed to 3 neurons; Output layer: the 1-dimensional 3x1 vector feature of the full connection layer is mapped to a single value of 1x1 through linear transformation.

3. An ultra-short term photovoltaic power output prediction system based on data parallel processing, characterized in that, The method comprises the following steps: The weather prediction data acquisition unit is configured to acquire weather prediction data in a first preset time period. The parallel processing unit is configured to perform the following three steps in parallel through three independent processes: Step S2-1: based on the weather prediction data, historical similar weather data is extracted from historical weather data, and corresponding historical photovoltaic output data is acquired based on the historical similar weather data; the XGBoost model is trained by taking the historical similar weather data and the historical photovoltaic output data as a training set; the weather prediction data is input into the trained XGBoost model to obtain statistical photovoltaic output prediction data; Step S2-2: according to the photovoltaic panel output characteristics and the weather prediction data, mechanism photovoltaic output prediction data is obtained; Step S2-3: the statistical photovoltaic output prediction data, the mechanism photovoltaic output prediction data, and the historical weather data and the corresponding historical photovoltaic output data in the second preset time period are combined into an input data set, a convolutional neural network is trained, and final predicted photovoltaic output data is generated; The first preset time period is 4 hours from the prediction time point; The weather prediction data includes weather type, wind speed, wind direction, temperature, relative humidity, visibility, air pressure, irradiance, sunrise and sunset time; In step S2-2, the calculation formula of the mechanism photovoltaic output prediction data is as follows: P PV = rgη max A PV η PV cosθ where P PV is the mechanistic photovoltaic power prediction data, r is the irradiance, g is the sunrise sunset encoding value, η max is the maximum power point, A PV is the photovoltaic radiation area, η PV is the conversion factor, and θ is the photovoltaic incidence angle. Step S2-3, specifically includes the following: Step S2-3-1: sequentially performing the following operations for each prediction time point in the prediction time domain, forming a plurality of sub-data sets, and merging the plurality of sub-data sets into an input data set: selecting statistical photovoltaic output prediction data, mechanism photovoltaic output prediction data, and historical meteorological data and corresponding historical photovoltaic output data in a second preset time period of the same time points of a preset number of historical similar weather days for sequence splicing, each time point containing 10 weather feature attributes, and after splicing, an original feature matrix with a dimension of 200x10 is formed as input data of a convolutional neural network input layer, the 10 weather attributes include a time stamp accurate to seconds, an encoded weather type, a wind speed, a wind direction encoded value, a temperature, a relative humidity, a visibility, a pressure, an irradiance, and a sunrise and sunset encoded value; one output attribute constitutes a 1x1 output vector, representing the historical photovoltaic output data of the corresponding time point, as the result target value of the parameter correction of the convolutional neural network; Step S2-3-2: training the convolutional neural network based on the input data set to obtain a final prediction result composed of prediction values of a plurality of prediction time points; The preset number is 200, and the second preset time period is 4 hours before the prediction time point.

Images