Short-term power prediction method, device and equipment of photovoltaic power station, medium and product

Abstract

The invention discloses a short-term power prediction method, device and equipment for a photovoltaic power station, a medium and a product. The method comprises the following steps: acquiring historical actual environment data and historical actual power data of multiple sources, and extracting a feature vector of each data source; an attention mechanism is adopted to adaptively distribute the corresponding feature weight of each data source according to the importance degree of the environment variable; performing weighted summation on the feature vectors according to the feature weights to generate fused feature vectors, and performing feature selection based on the fused feature vectors to obtain an optimized feature set; training a hybrid prediction model combining a classification algorithm and a time series data prediction algorithm according to the optimized feature set; the classification algorithm is used for judging a weather mode, and the time series data prediction algorithm is used for performing power prediction based on the weather mode; and acquiring environment prediction data, and generating a power prediction result by using the trained hybrid prediction model. Therefore, more accurate and stable power prediction is provided for nonlinear offshore data with high-dimensional and high-frequency fluctuations.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of photovoltaic power stations, and in particular to a short-term power prediction method, device, equipment, medium and product of a photovoltaic power station. Background Art

[0002] For photovoltaic power plants, especially floating photovoltaic power plants at sea, short-term power forecasting is very complex. Floating photovoltaic power plants at sea are affected by a variety of dynamic and complex environmental factors, such as frequently changing meteorological conditions (such as cloud cover, wind speed fluctuations, humidity and temperature changes) and unique ocean influences (such as waves, tides, seawater reflections, etc.), which have a significant impact on the power output of photovoltaic power plants. At the same time, the high noise, nonlinearity and uncertainty of offshore data also make it very difficult to deal with short-term power forecasting problems. Traditional forecasting schemes show deficiencies in processing efficiency and prediction accuracy when faced with high-frequency volatility and multi-dimensional characteristics of environmental parameters. Summary of the invention

[0003] The embodiments of the present invention provide a short-term power prediction method, device, equipment, medium and product for a photovoltaic power station, so as to achieve efficient and accurate prediction of the short-term power output of a photovoltaic power station under complex environmental conditions.

[0004] In a first aspect, an embodiment of the present invention provides a short-term power prediction method for a photovoltaic power station, the method comprising:

[0005] Acquire historical actual environmental data and historical actual power data from multiple sources, and extract feature vectors from each data source;

[0006] The attention mechanism is used to adaptively assign feature weights corresponding to each data source according to the importance of environmental variables;

[0007] Performing weighted summation on the feature vectors according to the feature weights to generate a fused feature vector, and performing feature selection based on the fused feature vector to obtain an optimized feature set;

[0008] Training a hybrid prediction model combining a classification algorithm and a time series data prediction algorithm according to the optimized feature set; the classification algorithm is used to determine weather patterns, and the time series data prediction algorithm is used to perform power prediction based on the weather patterns;

[0009] Acquire environmental prediction data, and use the trained hybrid prediction model to generate power prediction results.

[0010] Optionally, after acquiring the environmental prediction data and using the trained hybrid prediction model to generate a power prediction result, the method further includes:

[0011] Acquire actual power data corresponding to the power prediction result, and determine a prediction error based on the power prediction result and the actual power data;

[0012] The hybrid prediction model is optimized based on the prediction error, the target optimization index is monitored in real time during the optimization process, and the particle swarm optimization algorithm, the Bayesian optimization algorithm and the genetic algorithm are switched in turn according to the target optimization index.

[0013] Optionally, after optimizing the hybrid prediction model based on the prediction error, the method further includes:

[0014] The reinforcement learning mechanism is used to learn according to the historical optimization strategy selected in each optimization process to optimize the future optimization strategy.

[0015] Optionally, before training the hybrid prediction model combining the classification algorithm and the time series data prediction algorithm according to the optimized feature set, the method further includes:

[0016] A convolutional neural network is used to extract the time step features of the optimized feature set and input them into a recurrent neural network for time series modeling. During the modeling process, an attention mechanism is used to weight the corresponding local features according to the context of the current time step to determine the outliers in the optimized feature set;

[0017] The abnormal values ​​are corrected based on historical similar data.

[0018] Optionally, the hybrid prediction model includes a plurality of sub-models; and before training the hybrid prediction model combining the classification algorithm and the time series data prediction algorithm according to the optimized feature set, it also includes:

[0019] Using a generative adversarial network to perform data augmentation on the optimized feature set and perform classification based on weather patterns;

[0020] Accordingly, the hybrid prediction model combining the classification algorithm and the time series data prediction algorithm is trained according to the optimized feature set, including:

[0021] The sub-models corresponding to the corresponding weather patterns are trained respectively according to the classified data.

[0022] Optionally, the using the trained hybrid prediction model to generate a power prediction result includes:

[0023] The computing task of generating the power prediction result is distributed to a plurality of distributed computing nodes for parallel processing to obtain the power prediction result.

[0024] In a second aspect, an embodiment of the present invention further provides a short-term power prediction device for a photovoltaic power station, the device comprising:

[0025] A source data acquisition module is used to acquire historical actual environmental data and historical actual power data from multiple sources, and extract feature vectors from each data source respectively;

[0026] The feature weight allocation module is used to adaptively allocate the feature weights corresponding to each data source according to the importance of environmental variables using the attention mechanism;

[0027] A feature set optimization module, used for performing weighted summation on the feature vectors according to the feature weights to generate a fused feature vector, and performing feature selection based on the fused feature vector to obtain an optimized feature set;

[0028] A prediction model training module, which trains a hybrid prediction model combining a classification algorithm and a time series data prediction algorithm according to the optimized feature set; the classification algorithm is used to determine the weather pattern, and the time series data prediction algorithm is used to perform power prediction based on the weather pattern;

[0029] The prediction result generation module is used to obtain environmental prediction data and generate power prediction results using the trained hybrid prediction model.

[0030] In a third aspect, an embodiment of the present invention further provides a computer device, the computer device comprising:

[0031] one or more processors;

[0032] A memory for storing one or more programs;

[0033] When the one or more programs are executed by the one or more processors, the one or more processors implement the short-term power prediction method for a photovoltaic power station provided by any embodiment of the present invention.

[0034] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the short-term power prediction method for a photovoltaic power station provided by any embodiment of the present invention.

[0035] In a fifth aspect, an embodiment of the present invention further provides a computer program product, which includes a computer program. When the program is executed by a processor, the short-term power prediction method for a photovoltaic power station provided by any embodiment of the present invention is implemented.

[0036] The embodiment of the present invention provides a short-term power prediction method for a photovoltaic power station. First, historical actual environmental data and historical actual power data from multiple sources are obtained, and feature vectors of each data source are extracted respectively. Then, the feature weights corresponding to each data source are adaptively allocated according to the importance of environmental variables using an attention mechanism, and the feature vectors of each data source are weighted and summed according to each feature weight to generate a fused feature vector. Feature selection is then performed based on the fused feature vector to obtain an optimized feature set. Then, a hybrid prediction model combining a classification algorithm and a time series data prediction algorithm is trained according to the obtained optimized feature set, wherein the classification algorithm is used to determine the weather mode, and the time series data prediction algorithm is used to perform power prediction based on the obtained weather mode. Subsequently, future environmental prediction data can be obtained, and the trained hybrid prediction model is used to generate a power prediction result. The short-term power prediction method for a photovoltaic power station provided by the embodiment of the present invention combines the classification algorithm and the time series data prediction algorithm. The classification algorithm can fully mine the local features in a complex environment, and the time series data prediction algorithm can capture the deep relationship and nonlinear features between environmental factors and power output, so that a more accurate and stable power prediction can be provided when facing high-dimensional, high-frequency fluctuating nonlinear offshore data. At the same time, through the proposed multi-source data fusion process based on the attention mechanism, the model can adaptively adjust the weights of each data source according to the changes in complex environmental characteristics, so as to better capture the impact of environmental factors on power fluctuations, so that the model has higher prediction accuracy and robustness in complex and changing environments, making the short-term power prediction of photovoltaic power stations more efficient and accurate, and thus providing a reliable basis for actual grid scheduling and energy management. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A flowchart of a short-term power prediction method for a photovoltaic power station provided in Embodiment 1 of the present invention;

[0038] Figure 2 A schematic diagram of the structure of a short-term power prediction device for a photovoltaic power station provided in Embodiment 2 of the present invention;

[0039] Figure 3 This is a schematic diagram of the structure of a computer device provided in Embodiment 3 of the present invention. DETAILED DESCRIPTION

[0040] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

[0041] It should be mentioned before discussing the exemplary embodiments in more detail that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe the steps as sequential processes, many of the steps therein can be implemented in parallel, concurrently or simultaneously. In addition, the order of the steps can be rearranged. The process can be terminated when its operation is completed, but can also have additional steps not included in the accompanying drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0042] Embodiment 1

[0043] Figure 1 This is a flow chart of a method for short-term power prediction of a photovoltaic power station provided in Embodiment 1 of the present invention. This embodiment is applicable to the case of predicting the short-term power of a photovoltaic power station, especially an offshore floating photovoltaic power station. The method can be executed by a short-term power prediction device for a photovoltaic power station provided in an embodiment of the present invention, which can be implemented in hardware and / or software and can generally be integrated into a computer device. Figure 1 As shown, the specific steps include:

[0044] S11. Acquire historical actual environmental data and historical actual power data from multiple sources, and extract feature vectors from each data source.

[0045] S12. Use the attention mechanism to adaptively assign feature weights corresponding to each data source according to the importance of environmental variables.

[0046] S13. Performing weighted summation on the feature vectors according to the feature weights to generate a fused feature vector, and performing feature selection based on the fused feature vector to obtain an optimized feature set.

[0047] S14. Training a hybrid prediction model combining a classification algorithm and a time series data prediction algorithm according to the optimized feature set; the classification algorithm is used to determine weather patterns, and the time series data prediction algorithm is used to perform power prediction based on the weather patterns.

[0048] S15. Acquire environmental prediction data, and use the trained hybrid prediction model to generate a power prediction result.

[0049] An offshore floating photovoltaic power station is a photovoltaic power generation system installed on the water surface (such as the ocean or lake). The photovoltaic modules are fixed on the water surface by a floating device. By utilizing the water surface space, the land occupation is effectively reduced, and the cooling effect can be achieved, which can improve the efficiency of the photovoltaic panels. Short-term power forecasting refers to predicting the power output of the photovoltaic power station in the next few minutes to hours by analyzing historical data, which is crucial for the load management and power dispatching of the power grid. Taking an offshore floating photovoltaic power station as an example, historical actual environmental data and historical actual power data from multiple sources can be collected from meteorological forecasting systems, meteorological measurement equipment, sensors and power monitoring equipment of photovoltaic power stations, etc. Among them, the historical actual environmental data can include meteorological data (such as temperature, wind speed, irradiance, etc.) and ocean data (such as wave height, tide, current, etc.), and the time resolution of collecting multi-source data can be 15 minutes. In order to ensure the timeliness and accuracy of the data, a data acquisition gateway can be used for real-time transmission so that the data can reach the processing center smoothly.

[0050] After obtaining the historical actual environmental data and historical actual power data, the data can optionally be preprocessed, such as complete, normalized, and preliminary anomaly detection. Specifically, the interpolation method can be used to fill the missing values ​​in the data to complete the data with good time continuity. For data with close multi-dimensional relationships, a missing value filling method based on the K nearest neighbor algorithm (KNN) can be used to ensure the integrity of the data. Normalization technology can be used to convert data of different dimensions to the same scale. Commonly used methods include Min-Max normalization and Z-score standardization to eliminate the impact caused by dimensional differences between features. Statistical methods (such as Z-score, IQR) and machine learning methods (such as the isolation forest algorithm) can be combined to detect and clean outliers in the data to ensure the rationality of the input data.

[0051] Subsequently, multi-source data fusion can be performed to integrate data from multiple sources (such as meteorological data, ocean data, and historical actual power data) to improve the comprehensiveness and accuracy of the prediction. Specifically, a multi-layer perceptron (MLP) can be used to extract features for each data source to obtain different feature vectors. Then, the attention mechanism can be used to calculate the feature weights of each data source, and the weights are adaptively allocated according to the importance of the environmental variables to ensure that the most important data features receive higher feature weights during the fusion process. The feature vectors from different sources are then weighted and summed according to the feature weights to generate the final fused feature vector. Compared with traditional fusion methods such as simple splicing or weighted averaging, by introducing the attention mechanism, the weights of different data sources can be automatically judged and allocated according to the dynamic changes of environmental characteristics to highlight the variables that have the greatest impact on power fluctuations, thereby improving the model's ability to handle complex data, as well as the accuracy and robustness of power prediction.

[0052] After completing the fusion of multi-source data, feature selection can be performed to reduce the dimension of the original data and automatically select the most valuable features for prediction, thereby improving the efficiency and prediction accuracy of the model. Specifically, the importance of each feature can be evaluated through Bayesian optimization or recursive feature elimination (RFE) method, and then the most representative features can be selected based on the evaluation results, and low-importance features can be eliminated to obtain the final optimized feature set.

[0053] The hybrid prediction model can combine the classification algorithm and the time series data prediction algorithm to form an end-to-end hybrid model framework, so as to take advantage of the advantages of different algorithms and improve the prediction accuracy. Among them, the classification algorithm is responsible for the preliminary pattern recognition of the input data, while the time series data prediction algorithm performs specific power prediction for each pattern, thereby improving the accuracy and adaptability of the overall model. The classification algorithm is preferably the K nearest neighbor algorithm (KNN), which can identify different weather patterns, such as sunny days, cloudy days, rainy days, etc., so as to better handle the power prediction in the marine environment. The time series data prediction algorithm can use a neural network (NN), which can perform power prediction under a certain weather pattern. It is preferably a long short-term memory network (LSTM), which is good at processing and predicting long-distance dependencies in time series data, can capture the time-dependent characteristics of power fluctuations in the marine environment, and is suitable for processing complex power fluctuations. Of course, a multi-layer perceptron (MLP) can also be used to process nonlinear features and is suitable for processing simpler weather patterns and power predictions. After obtaining the optimized feature set, the hybrid prediction model can be trained using the optimized feature set.

[0054] Then, the environmental prediction data required for power prediction can be obtained, and the environmental prediction data can be processed accordingly, so as to use the trained hybrid prediction model to predict the short-term power prediction results. Corresponding to the time resolution of collecting multi-source data, the hybrid prediction model can output the power prediction results every 15 minutes and transmit them to the control center or other decision-making systems for power grid scheduling and photovoltaic power station power management. Accordingly, the control center or other decision-making systems can display the power prediction results through visualization tools for operators to understand and analyze.

[0055] On the basis of the above technical solution, optionally, after obtaining the environmental prediction data and using the trained hybrid prediction model to generate a power prediction result, it also includes: obtaining actual power data corresponding to the power prediction result, and determining a prediction error based on the power prediction result and the actual power data; optimizing the hybrid prediction model based on the prediction error, monitoring the target optimization index in real time during the optimization process, and switching between the particle swarm optimization algorithm, the Bayesian optimization algorithm and the genetic algorithm in turn according to the target optimization index.

[0056] Specifically, after each prediction is completed, the error analysis method (such as mean square error, absolute error) can be used to determine the prediction error between the power prediction result and the actual power data at the corresponding time point to evaluate the performance of the model. At the same time, the hybrid prediction model can be adaptively optimized based on the prediction error to form an adaptive feedback loop, thereby further improving the accuracy and robustness of the prediction. The feedback loop takes the real-time performance of the model as the core, continuously updates the optimization strategy, and can feed back the results of each strategy selection (such as the effect of reducing the loss function) to the system to form the historical data of the optimization strategy. The adaptive hybrid optimization algorithm combines the particle swarm optimization algorithm (PSO), the Bayesian optimization algorithm and the genetic algorithm (GA), and can dynamically select the optimal model parameter adjustment method according to the changes in the real-time environment. Among them, the particle swarm optimization algorithm is used to search for the optimal solution globally, the genetic algorithm is used to enhance the robustness of the model, and the Bayesian optimization is used to efficiently find the best hyperparameter combination in a complex search space. In each training cycle, the optimization algorithm adjusts the model parameters according to the prediction error of the model to ensure rapid convergence and high-precision prediction of the model. Specifically, the particle swarm optimization algorithm can be used to perform global optimization of the initial weight parameters. The adjustment of the weight parameters is based on the prediction performance of the current model, thereby ensuring the best performance under different weather patterns and environmental conditions. Then the Bayesian optimization algorithm is used to fine-grainedly adjust the hyperparameters (such as learning rate, number of neurons, number of neighbors of KNN, etc.) to accelerate model convergence and reduce overfitting. Finally, the generalization ability of the model is further enhanced by the genetic algorithm. Specifically, the number of layers and structures of the model can be adaptively adjusted according to the complexity of the data and the changes in the real-time environment to ensure that the optimal computing resource utilization can be achieved in different situations, thereby reducing the training and running time of the model, and ensuring that efficient power prediction can be provided in real time at a time resolution of 15 minutes. Exemplarily, for weather patterns with higher complexity, the number of layers of the neural network can be increased to enhance the expressive power of the model, while for simpler cases, the number of layers can be reduced to improve the computational efficiency of the model. Compared with traditional optimization schemes, by combining an adaptive hybrid optimization algorithm, the hybrid prediction model can dynamically adapt to changes in the environment and data and respond quickly. In particular, the adaptive structural adjustment characteristics of the genetic algorithm enable the model to be dynamically reconstructed according to real-time changes in the environment, avoiding model failure in the face of extreme weather conditions or sudden environmental changes, achieving a balance between resource efficiency and prediction accuracy, and being able to maintain good prediction performance when faced with complex time series data, which has strong application value in marine environment prediction scenarios.

[0057] The adaptive hybrid optimization algorithm adapts to the complex and changeable environmental characteristics of offshore photovoltaic power prediction tasks by dynamically adjusting and selecting different optimization strategies. Particle swarm optimization algorithm, Bayesian optimization algorithm and genetic algorithm are combined in a modular way, using the characteristics of each optimization algorithm to form an overall optimization framework, which plays a role in different stages to continuously improve the adaptability and optimization efficiency of the model. The entire optimization process is divided into three stages: initial global search, local fine optimization and adaptive adjustment. Different optimization strategies are used in different stages to form a collaborative optimization closed loop. Particle swarm optimization algorithm, Bayesian optimization algorithm and genetic algorithm are not simple linear combinations, but modular collaboration according to the staged requirements of optimization, so that the entire optimization process is adaptive and staged, and each algorithm plays the greatest role in the stage where it is best. The adaptive hybrid optimization algorithm has the ability to switch dynamically, and can select the appropriate optimization strategy according to the performance of the model at different stages. For example, when the Bayesian optimization algorithm cannot significantly reduce the loss function, it can automatically switch to the genetic algorithm for parameter reorganization and mutation to jump out of the local optimum.

[0058] For the dynamic switching process, the target optimization index can be monitored in real time during the optimization process. The target optimization index can include the convergence speed of the model, the rate of change of the loss function, the stability of the parameters, the utilization rate of computing resources, etc. These indicators jointly determine the state of the current optimization stage and whether the optimization strategy needs to be adjusted. Further, for different target optimization indicators, specific thresholds can be set to determine whether switching is needed. For example, when the convergence speed is lower than a certain set threshold, or the loss function does not drop significantly after multiple iterations, the optimization strategy switching can be started. The switching process is specifically described. In the initial stage, the particle swarm optimization algorithm can be used for global search. The particle swarm optimization algorithm is suitable for finding multiple possible better solutions in the parameter space to give the model a good starting point. When the particle swarm optimization algorithm finds a better parameter combination, when it is detected that the rate of change of the loss function tends to be stable and the model enters the fine-tuning stage, it can be dynamically switched to the Bayesian optimization algorithm. When the model enters a relatively stable state but there is an optimization bottleneck, or when the data environment changes significantly (for example, the environmental data fluctuates greatly and the characteristic statistical distribution changes), the genetic algorithm can be started for adaptive adjustment.

[0059] Further optionally, after optimizing the hybrid prediction model based on the prediction error, it also includes: using a reinforcement learning mechanism to learn according to the historical optimization strategy selected in each optimization process to optimize the future optimization strategy.

[0060] Specifically, a reinforcement learning mechanism can be used to track the performance of each selection of different optimization strategies during the long-term optimization process, and learn the optimal strategy combination through experiments and feedback, so that the model's selection strategy can gradually evolve and adapt to different optimization tasks and data environments. For example, if in some scenarios, the effect of switching to genetic algorithm optimization is better, the system will gradually prefer this choice, thereby gradually optimizing the entire strategy selection process.

[0061] On the basis of the above technical solution, optionally, before training the hybrid prediction model combining the classification algorithm and the time series data prediction algorithm according to the optimized feature set, it also includes: using a convolutional neural network to extract the time step features of the optimized feature set and inputting them into a recurrent neural network for time series modeling, and using an attention mechanism in the modeling process to weight the corresponding local features according to the context of the current time step to determine the outliers in the optimized feature set; and correcting the outliers according to historical similar data.

[0062] Specifically, before each training, deep learning technology can be used to detect and clean the outliers in the optimized feature set obtained by combining convolutional neural network (CNN) and recurrent neural network (RNN) to filter the noise in meteorological and ocean data, thereby improving the quality of data and the prediction accuracy of the model, effectively reducing the interference of noise on model prediction, and making the model more robust and stable when facing complex marine environments. Among them, the convolutional neural network is used to extract local features from time segments, and the recurrent neural network is used to capture the dependencies of time series. The convolutional neural network and the recurrent neural network are combined in a cascade manner. First, the convolutional neural network extracts the features of each time step of the optimized feature set, and then these features are input into the recurrent neural network for time series modeling. Compared with the traditional cleaning scheme based on simple screening of statistical characteristics, this cascade structure can not only effectively identify local anomalies, but also combine time dependency to further improve the accuracy of noise detection, so as to ensure the comprehensive analysis of local noise and global trends, so as to more accurately identify noise. When a recurrent neural network is used to model time series, an attention mechanism can be introduced to weight the corresponding local features extracted by the convolutional neural network according to the context of the current time step to highlight features that may be noise, so that the model can identify abnormal data points more flexibly and accurately.

[0063] After the outliers are determined, they can be corrected according to historical similar data. Specifically, the outliers can be interpolated by using the normal time series of historical similar data to replace the outliers by interpolation or prediction. For example, for the wind speed data of a certain day, if it is detected as abnormal, the system will refer to the wind speed data under similar meteorological conditions in history for interpolation, and specifically linear interpolation, average value based on adjacent data points, etc. can be used. Recurrent neural networks can also be used to predict the trend of the data of the previous and next time steps. When a data point is detected to be abnormal, the recurrent neural network will combine the past sequence information and the future time step trend, and use its predicted results to replace the outliers. Specifically, LSTM can be used to perform time dependency analysis on past and future data, and the reasonable value of the current outlier point can be predicted in combination with the trend of the entire time window. Further, in order to increase the robustness of the correction, multiple models can be combined, such as the above two methods, and the output results of multiple models can be fused, such as weighted average or majority voting, to finally determine the corrected data value. Such a multi-source correction mechanism can show extremely high adaptability in dynamic and complex marine environments and meteorological data. In addition, different correction strategies can be selected according to the different types of noise. Exemplarily, for random noise, smoothing filtering (such as moving average) can be used to reduce the impact of noise, while for mutational noise, it can be corrected by relying more on the trend prediction of recurrent neural networks. The identification of noise types can be achieved during the feature extraction process of convolutional neural networks. Noise types can include periodic anomalies, mutational anomalies, etc., so that appropriate correction strategies can be selected according to the obtained noise types. The ability to classify and process different types of noise makes the system more accurate in processing complex and diverse noises. Through the above-mentioned outlier identification and automatic correction, there is no need for a large amount of manual intervention, and its automation and real-time nature are particularly suitable for marine environments that require quick decision-making.

[0064] Based on the above technical solution, optionally, the hybrid prediction model includes multiple sub-models; before the hybrid prediction model combining the classification algorithm and the time series data prediction algorithm is trained according to the optimized feature set, it also includes: using a generative adversarial network to perform data enhancement on the optimized feature set, and classifying based on weather patterns; accordingly, the hybrid prediction model combining the classification algorithm and the time series data prediction algorithm is trained according to the optimized feature set, including: training the sub-models corresponding to the corresponding weather patterns according to the classified data.

[0065] Specifically, the generative adversarial network (GAN) consists of a generator and a discriminator. The generator generates virtual samples similar to real data, while the discriminator is used to distinguish between false and real data. The two are constantly improved in the confrontation, and finally generate synthetic data with high quality and similar distribution to real data. The obtained optimized feature set can be adaptively enhanced by the generative adversarial network to enrich the data diversity. Specifically, a dynamic data enhancement strategy can be adopted to adjust the number and type of enhanced samples according to real-time data features. Exemplarily, when the data is insufficient at the beginning of training or when the system detects data imbalance, data samples of certain features will be automatically added to ensure that the training process of the model is more comprehensive and improve the generalization ability of the prediction model. The data generated by the generative adversarial network can cover photovoltaic power output under different weather modes (such as sunny days, cloudy days, strong winds, etc.), and the enhanced data can be further automatically classified to obtain training sample subsets corresponding to different weather modes, so that the sub-models under the corresponding weather modes can be trained according to each training sample subset. By constructing special prediction models for different weather modes, better adaptability and accuracy can be ensured in complex marine environments. Accordingly, during the prediction process, the current weather pattern can be identified through a classification algorithm (such as the clustering-based K-means algorithm), and then the sub-model that best suits the current weather pattern can be automatically selected for prediction based on the identification result, thereby ensuring optimal prediction accuracy under different environments.

[0066] For the time series data prediction algorithm in the hybrid prediction model, a multi-model integration strategy can also be adopted, such as integrating long short-term memory network (LSTM), gated recurrent unit (GRU) and gradient boosting tree model (XGBoost). The gated recurrent unit is similar to the long short-term memory network. It is a simplified recurrent neural network that reduces the number of parameters and computational complexity, and has good performance in predicting short-term power. The gradient boosting tree model is efficient, flexible, and accurate, and is suitable for processing complex regression and classification tasks. Each model has different advantages. LSTM and GRU are good at processing long-term and short-term dependencies of time series data, while XGBoost has unique advantages in processing feature importance and nonlinear relationships. By integrating multiple models, the prediction accuracy and robustness can be improved, which is particularly suitable for complex and changeable offshore photovoltaic power prediction scenarios. Furthermore, an adaptive weight allocation mechanism based on prediction error feedback can be adopted. In the real-time prediction process, the weights of each model are dynamically adjusted according to its historical performance, and the output results of each model are weighted fused. This allows the most effective model combination to be used under different weather patterns, thereby achieving both the complex feature fitting capabilities of a deep learning model and the stability and speed of a traditional machine learning model.

[0067] On the basis of the above technical solution, optionally, the use of the trained hybrid prediction model to generate a power prediction result includes: assigning the computing task of generating the power prediction result to multiple distributed computing nodes for parallel processing to obtain the power prediction result. By distributing the computing tasks to multiple distributed computing nodes to achieve parallel processing, the computing efficiency and real-time performance of the model are improved, especially when processing large amounts of data. Among them, the intermediate processing results are shared between the computing nodes. Furthermore, it is also possible to combine edge computing to assign computing tasks to locations as close to the corresponding data sources as possible, thereby reducing the delay in data transmission, further improving the overall computing efficiency, and enabling the power prediction model to quickly respond to high-frequency data changes in complex marine environments, ensuring real-time and accuracy.

[0068] The technical solution provided by the embodiment of the present invention first obtains historical actual environmental data and historical actual power data from multiple sources, and extracts the feature vectors of each data source respectively, then uses the attention mechanism to adaptively allocate the feature weights corresponding to each data source according to the importance of the environmental variables, and performs weighted summation of the feature vectors of each data source according to each feature weight to generate a fused feature vector, and then performs feature selection based on the fused feature vector to obtain an optimized feature set, and then trains a hybrid prediction model combining a classification algorithm and a time series data prediction algorithm according to the obtained optimized feature set, wherein the classification algorithm is used to determine the weather pattern, and the time series data prediction algorithm is used to perform power prediction based on the obtained weather pattern, and then the future environmental prediction data can be obtained, and the trained hybrid prediction model is used to generate a power prediction result. By combining the classification algorithm and the time series data prediction algorithm, the classification algorithm can fully explore the local features in a complex environment, and the time series data prediction algorithm can capture the deep relationship and nonlinear characteristics between environmental factors and power output, so that it can provide more accurate and stable power prediction when facing high-dimensional, high-frequency fluctuating nonlinear offshore data. At the same time, through the proposed multi-source data fusion process based on the attention mechanism, the model can adaptively adjust the weights of each data source according to the changes in complex environmental characteristics, so as to better capture the impact of environmental factors on power fluctuations, so that the model has higher prediction accuracy and robustness in complex and changing environments, making the short-term power prediction of photovoltaic power stations more efficient and accurate, and thus providing a reliable basis for actual grid scheduling and energy management.

[0069] Embodiment 2

[0070] Figure 2 This is a schematic diagram of the structure of a short-term power prediction device for a photovoltaic power station provided in the second embodiment of the present invention. The device can be implemented by hardware and / or software, and can generally be integrated into a computer device to execute the short-term power prediction method for a photovoltaic power station provided in any embodiment of the present invention. Figure 2 As shown, the device comprises:

[0071] The source data acquisition module 21 is used to acquire historical actual environmental data and historical actual power data from multiple sources, and extract the feature vectors of each data source respectively;

[0072] A feature weight allocation module 22, used to adaptively allocate feature weights corresponding to each data source according to the importance of environmental variables using an attention mechanism;

[0073] A feature set optimization module 23, configured to perform weighted summation on the feature vectors according to the feature weights to generate a fused feature vector, and perform feature selection based on the fused feature vector to obtain an optimized feature set;

[0074] A prediction model training module 24 is used to train a hybrid prediction model combining a classification algorithm and a time series data prediction algorithm according to the optimized feature set; the classification algorithm is used to determine the weather pattern, and the time series data prediction algorithm is used to perform power prediction based on the weather pattern;

[0075] The prediction result generating module 25 is used to obtain environmental prediction data and generate power prediction results using the trained hybrid prediction model.

[0076] The technical solution provided by the embodiment of the present invention first obtains historical actual environmental data and historical actual power data from multiple sources, and extracts the feature vectors of each data source respectively, then uses the attention mechanism to adaptively allocate the feature weights corresponding to each data source according to the importance of the environmental variables, and performs weighted summation of the feature vectors of each data source according to each feature weight to generate a fused feature vector, and then performs feature selection based on the fused feature vector to obtain an optimized feature set, and then trains a hybrid prediction model combining a classification algorithm and a time series data prediction algorithm according to the obtained optimized feature set, wherein the classification algorithm is used to determine the weather pattern, and the time series data prediction algorithm is used to perform power prediction based on the obtained weather pattern, and then the future environmental prediction data can be obtained, and the trained hybrid prediction model is used to generate a power prediction result. By combining the classification algorithm and the time series data prediction algorithm, the classification algorithm can fully explore the local features in a complex environment, and the time series data prediction algorithm can capture the deep relationship and nonlinear characteristics between environmental factors and power output, so that it can provide more accurate and stable power prediction when facing high-dimensional, high-frequency fluctuating nonlinear offshore data. At the same time, through the proposed multi-source data fusion process based on the attention mechanism, the model can adaptively adjust the weights of each data source according to the changes in complex environmental characteristics, so as to better capture the impact of environmental factors on power fluctuations, so that the model has higher prediction accuracy and robustness in complex and changing environments, making the short-term power prediction of photovoltaic power stations more efficient and accurate, and thus providing a reliable basis for actual grid scheduling and energy management.

[0077] On the basis of the above technical solution, optionally, the device further includes:

[0078] A prediction error determination module, configured to, after acquiring the environmental prediction data and generating a power prediction result using the trained hybrid prediction model, obtain actual power data corresponding to the power prediction result, and determine a prediction error based on the power prediction result and the actual power data;

[0079] The prediction model optimization module is used to optimize the hybrid prediction model based on the prediction error. The optimization process monitors the target optimization index in real time and switches between the particle swarm optimization algorithm, the Bayesian optimization algorithm and the genetic algorithm in turn according to the target optimization index.

[0080] On the basis of the above technical solution, optionally, the device further includes:

[0081] The optimization strategy optimization module is used to use a reinforcement learning mechanism to learn according to the historical optimization strategy selected in each optimization process after optimizing the hybrid prediction model based on the prediction error, so as to optimize the future optimization strategy.

[0082] On the basis of the above technical solution, optionally, the device further includes:

[0083] An outlier determination module is used to use a convolutional neural network to extract the time step features of the optimized feature set and input them into a recurrent neural network for time series modeling before training the hybrid prediction model combining the classification algorithm and the time series data prediction algorithm according to the optimized feature set. During the modeling process, an attention mechanism is used to weight the corresponding local features according to the context of the current time step to determine the outliers in the optimized feature set;

[0084] The outlier correction module is used to correct the outlier based on historical similar data.

[0085] On the basis of the above technical solution, optionally, the hybrid prediction model includes a plurality of sub-models; the device further includes:

[0086] A data enhancement module, used for performing data enhancement on the optimized feature set using a generative adversarial network and performing classification based on weather patterns before training the hybrid prediction model combining the classification algorithm and the time series data prediction algorithm according to the optimized feature set;

[0087] Accordingly, the prediction model training module 24 is specifically used for:

[0088] The sub-models corresponding to the corresponding weather patterns are trained respectively according to the classified data.

[0089] On the basis of the above technical solution, optionally, the prediction result generation module 25 is specifically used for:

[0090] The computing task of generating the power prediction result is distributed to a plurality of distributed computing nodes for parallel processing to obtain the power prediction result.

[0091] The short-term power prediction device for a photovoltaic power station provided by an embodiment of the present invention can execute the short-term power prediction method for a photovoltaic power station provided by any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0092] It is worth noting that in the embodiment of the short-term power prediction device of the above-mentioned photovoltaic power station, the various units and modules included are only divided according to functional logic, but are not limited to the above-mentioned division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.

[0093] Embodiment 3

[0094] Figure 3The schematic diagram of the structure of the computer device provided for the third embodiment of the present invention shows a block diagram of an exemplary computer device suitable for implementing the implementation mode of the present invention. Figure 3 The computer device shown is only an example and should not limit the functions and scope of use of the embodiments of the present invention. Figure 3 As shown, the computer device includes a processor 31, a memory 32, an input device 33 and an output device 34; the number of processors 31 in the computer device can be one or more. Figure 3 Taking a processor 31 as an example, the processor 31, memory 32, input device 33 and output device 34 in the computer device can be connected through a bus or other means. Figure 3 The example of connecting through bus is taken in the following.

[0095] The memory 32, as a computer-readable storage medium, can be used to store software programs, computer executable programs and modules, such as the program instructions / modules corresponding to the short-term power prediction method of the photovoltaic power station in the embodiment of the present invention (for example, the source data acquisition module 21, the feature weight allocation module 22, the feature set optimization module 23, the prediction model training module 24 and the prediction result generation module 25 in the short-term power prediction device of the photovoltaic power station). The processor 31 executes various functional applications and data processing of the computer device by running the software programs, instructions and modules stored in the memory 32, that is, realizes the above-mentioned short-term power prediction method of the photovoltaic power station.

[0096] The memory 32 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and at least one application required for a function; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 32 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 32 may further include a memory remotely arranged relative to the processor 31, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0097] The input device 33 can be used to obtain historical actual environmental data and historical actual power data from multiple sources in real time, and to generate key signal inputs related to user settings and function control of computer equipment, etc. The output device 34 can be used to transmit power prediction results to the control center, etc.

[0098] Embodiment 4

[0099] Embodiment 4 of the present invention further provides a storage medium containing computer executable instructions, which, when executed by a computer processor, is used to execute a short-term power prediction method for a photovoltaic power station, the method comprising:

[0100] Acquire historical actual environmental data and historical actual power data from multiple sources, and extract feature vectors from each data source;

[0101] The attention mechanism is used to adaptively assign feature weights corresponding to each data source according to the importance of environmental variables;

[0102] Performing weighted summation on the feature vectors according to the feature weights to generate a fused feature vector, and performing feature selection based on the fused feature vector to obtain an optimized feature set;

[0103] Training a hybrid prediction model combining a classification algorithm and a time series data prediction algorithm according to the optimized feature set; the classification algorithm is used to determine weather patterns, and the time series data prediction algorithm is used to perform power prediction based on the weather patterns;

[0104] Acquire environmental prediction data, and use the trained hybrid prediction model to generate power prediction results.

[0105] The storage medium may be any of various types of memory devices or storage devices. The term "storage medium" is intended to include: installation media, such as CD-ROM, floppy disk or tape device; computer system memory or random access memory, such as DRAM, DDR RAM, SRAM, EDO RAM, Rambus RAM, etc.; non-volatile memory, such as flash memory, magnetic media (such as hard disk or optical storage); registers or other similar types of memory elements, etc. The storage medium may also include other types of memory or combinations thereof. In addition, the storage medium may be located in the computer system in which the program is executed, or may be located in a different second computer system, which is connected to the computer system via a network (such as the Internet). The second computer system may provide program instructions to the computer for execution. The term "storage medium" may include two or more storage media that may reside in different locations (e.g., in different computer systems connected via a network). The storage medium may store program instructions (e.g., embodied as a computer program) that may be executed by one or more processors.

[0106] Of course, the computer executable instructions of a storage medium containing computer executable instructions provided in an embodiment of the present invention are not limited to the method operations described above, and can also execute related operations in the short-term power prediction method for a photovoltaic power station provided in any embodiment of the present invention.

[0107] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, which carry computer-readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0108] The program code embodied on the computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0109] Through the above description of the implementation methods, the technicians in the relevant field can clearly understand that the present invention can be implemented by means of software and necessary general hardware, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0110] Embodiment 5

[0111] Embodiment 5 of the present invention also provides a computer program product, which includes a computer program (also referred to as code, instruction), which can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it is used to execute the short-term power prediction method for a photovoltaic power station provided in any of the above embodiments, and has the corresponding beneficial effects of the execution method.

[0112] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A short-term power prediction method for a photovoltaic power station, characterized in that: include: Acquire historical actual environmental data and historical actual power data from multiple sources, and extract feature vectors from each data source; The attention mechanism is used to adaptively assign feature weights corresponding to each data source according to the importance of environmental variables; Performing weighted summation on the feature vectors according to the feature weights to generate a fused feature vector, and performing feature selection based on the fused feature vector to obtain an optimized feature set; Training a hybrid prediction model combining a classification algorithm and a time series data prediction algorithm according to the optimized feature set; the classification algorithm is used to determine weather patterns, and the time series data prediction algorithm is used to perform power prediction based on the weather patterns; Acquire environmental prediction data, and use the trained hybrid prediction model to generate power prediction results.

2. The short-term power prediction method of a photovoltaic power station according to claim 1, characterized in that: After acquiring the environmental prediction data and using the trained hybrid prediction model to generate a power prediction result, the method further includes: Acquire actual power data corresponding to the power prediction result, and determine a prediction error based on the power prediction result and the actual power data; The hybrid prediction model is optimized based on the prediction error, the target optimization index is monitored in real time during the optimization process, and the particle swarm optimization algorithm, the Bayesian optimization algorithm and the genetic algorithm are switched in turn according to the target optimization index.

3. The short-term power prediction method for a photovoltaic power station according to claim 2, characterized in that: After optimizing the hybrid prediction model based on the prediction error, the method further includes: The reinforcement learning mechanism is used to learn according to the historical optimization strategy selected in each optimization process to optimize the future optimization strategy.

4. The short-term power prediction method for a photovoltaic power station according to claim 1, characterized in that: Before training the hybrid prediction model combining the classification algorithm and the time series data prediction algorithm according to the optimized feature set, the method further includes: A convolutional neural network is used to extract the time step features of the optimized feature set and input them into a recurrent neural network for time series modeling. During the modeling process, an attention mechanism is used to weight the corresponding local features according to the context of the current time step to determine the outliers in the optimized feature set; The abnormal values ​​are corrected based on historical similar data.

5. The short-term power prediction method for a photovoltaic power station according to claim 1, characterized in that: The hybrid prediction model includes a plurality of sub-models; before the hybrid prediction model combining the classification algorithm and the time series data prediction algorithm is trained according to the optimized feature set, it also includes: Using a generative adversarial network to perform data augmentation on the optimized feature set and perform classification based on weather patterns; Accordingly, the hybrid prediction model combining the classification algorithm and the time series data prediction algorithm is trained according to the optimized feature set, including: The sub-models corresponding to the corresponding weather patterns are trained respectively according to the classified data.

6. The short-term power prediction method for a photovoltaic power station according to claim 1, characterized in that: The step of using the trained hybrid prediction model to generate a power prediction result comprises: The computing task of generating the power prediction result is distributed to a plurality of distributed computing nodes for parallel processing to obtain the power prediction result.

7. A short-term power prediction device for a photovoltaic power station, characterized in that: include: A source data acquisition module is used to acquire historical actual environmental data and historical actual power data from multiple sources, and extract feature vectors from each data source respectively; The feature weight allocation module is used to adaptively allocate the feature weights corresponding to each data source according to the importance of environmental variables using the attention mechanism; A feature set optimization module, used for performing weighted summation on the feature vectors according to the feature weights to generate a fused feature vector, and performing feature selection based on the fused feature vector to obtain an optimized feature set; A prediction model training module, which trains a hybrid prediction model combining a classification algorithm and a time series data prediction algorithm according to the optimized feature set; the classification algorithm is used to determine the weather pattern, and the time series data prediction algorithm is used to perform power prediction based on the weather pattern; The prediction result generation module is used to obtain environmental prediction data and generate power prediction results using the trained hybrid prediction model.

8. A computer device, characterized in that: include: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the short-term power prediction method for a photovoltaic power station as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the short-term power prediction method for a photovoltaic power station as claimed in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the computer program implements the short-term power prediction method for a photovoltaic power station as claimed in any one of claims 1 to 6.

Images