New energy photovoltaic power prediction system and method based on deep learning

Abstract

The invention relates to the technical field of deep learning, in particular to a new energy photovoltaic power prediction system and method based on deep learning. Comprising the steps of obtaining historical power data and meteorological data of a photovoltaic station, and performing power prediction of a current time sequence through a single-station power prediction model; building a multi-site joint prediction model to capture the spatial correlation between sites so as to predict the power of the photovoltaic sites; establishing a weather mode identification method according to the meteorological data to distinguish a normal weather mode and a sudden weather mode; and establishing an adversarial weather-power combined prediction model to process multi-site power collaborative prediction. According to the method, single-station power prediction and multi-station joint prediction are carried out through online learning and transfer learning, so that the station power prediction efficiency is improved; by simulating extreme and violently changing weather conditions, adversarial training is carried out to establish an adversarial weather prediction network, and then weather and power combined prediction is carried out, so that the accuracy of the prediction model in a severe environment is improved.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and specifically to a new energy photovoltaic power prediction system and method based on deep learning. Background Art

[0002] As the global energy transition process accelerates, the proportion of new energy, especially photovoltaic power generation, in the energy structure continues to increase; however, the nature of photovoltaic power generation determines that its output power has strong volatility and intermittency, which poses a significant challenge to the safe and stable operation of the power system. Therefore, photovoltaic power prediction is crucial for planning power plans in advance. Traditional photovoltaic power prediction methods mainly rely on simple statistical models, regression analysis or basic time series models; traditional methods usually use a single meteorological parameter and power linear relationship modeling, and the prediction error rate is generally high, especially under cloudy and sunny weather conditions, the error rate may exceed 40%; traditional method models are slow to respond to sudden weather changes, and most of them predict a single photovoltaic site, ignoring the spatial correlation between geographically adjacent sites; and traditional method models mostly use offline training methods, which are difficult to adapt to changes in system characteristics caused by changes in external factors. Summary of the invention

[0003] The purpose of the present invention is to provide a new energy photovoltaic power prediction system and method based on deep learning, so as to solve the problem of high error rate of site power prediction in single-site power prediction and multi-site joint prediction, and improve the power prediction efficiency; by simulating extreme and drastically changing weather conditions for adversarial training to establish an adversarial weather prediction network, and then perform joint prediction of weather and power, so as to improve the accuracy of the prediction model in harsh environments.

[0004] To achieve the above object, on the one hand, the present invention provides a new energy photovoltaic power prediction method based on deep learning, the method comprising the following steps: Step S1, obtaining historical power data and meteorological data of a photovoltaic site; the photovoltaic site is a single photovoltaic site or multiple photovoltaic sites within the research scope; the meteorological data includes temperature, humidity, wind speed and light intensity; and preprocessing the historical power data and meteorological data to obtain standardized data.

[0005] Step S2, extracting local pattern features of meteorological data and power data based on the standardized data, and performing power prediction for the current time series through a single-site power prediction model; the single-site power prediction model includes a preset time series prediction model and a preset spatial feature extraction model; and predicting photovoltaic site power by establishing a multi-site joint prediction model to capture the spatial correlation between sites.

[0006] Step S3: Establish a weather pattern recognition method based on the meteorological data to distinguish normal weather patterns and sudden weather patterns; establish an adversarial weather prediction network based on adversarial learning technology, where the adversarial weather prediction network is a deep learning framework that generates weather prediction data through adversarial training.

[0007] Step S4: Based on the adversarial weather prediction network, establish an adversarial weather-power joint prediction model, where the adversarial weather-power joint prediction model performs collaborative prediction of multi-site power through joint online learning and transfer learning of meteorological conditions and power output.

[0008] Furthermore, the preset spatial feature extraction model identifies and extracts local features of meteorological data and power data by learning the standardized data; the preset spatial feature extraction model uses a convolutional neural network for calculation, and the formula is: ; where is the value of the -th layer feature map at position ; is the value of the -th layer feature map at position ; is the weight of the -th layer convolution kernel at position ; is the height and width of the convolution kernel; is the configuration item of the -th layer; is a preset activation function; M is the number of rows of the convolution kernel, representing the number of pixels covered by the convolution kernel in the vertical direction; N is the number of columns of the convolution kernel, representing the number of pixels covered by the convolution kernel in the horizontal direction.

[0009] Furthermore, the method for power prediction includes: Capture the long-term and short-term temporal dependencies of the photovoltaic site by establishing the single-site power prediction model for power prediction at the current time series; first, receive the local feature values of meteorological data and power data as input features, and use the historical learning information as the hidden state; then, judge whether the hidden state information needs to be retained or discarded through the forget gate, screen the new information in the input features through the input gate and generate a candidate memory unit; update the memory unit by weighted fusion of the forgotten hidden state information and the new information; finally, generate the power prediction value of the photovoltaic site at the current moment according to the updated new memory unit.

[0010] Furthermore, the method for establishing the multi-site joint prediction model includes: The multi-site joint prediction model updates the power information of the photovoltaic site to be predicted by studying the spatial correlation between multiple sites. The formula is as follows: ; Wherein, is the feature representation of node at the th layer; represents the neighbor set of node ; represents the number of neighbors; represents the weight matrix of the th layer; is the configuration item of the th layer; is the feature representation of node at the th layer; is the preset activation function.

[0011] Furthermore, the weather pattern recognition method classifies different weather patterns by constructing a hyperplane, sets a preset decision boundary, and updates the hyperplane and then the preset decision boundary through data iteration; the normal weather pattern and the sudden weather pattern are distinguished by the preset decision boundary after iteration.

[0012] The adversarial training is a mutual game process between the generator and the discriminator: the generator continuously improves the quality of the generated weather data, and the discriminator improves the ability to distinguish real and generated data; the objective function of the adversarial weather prediction network is expressed as: ; Wherein, is the generator network; represents minimizing the generator error; is the discriminator network; represents maximizing the discriminator error; represents the objective functions of the generator and the discriminator; is the real data sample; is the distribution of the real weather data; is the expected value; represents is sampled from the real weather data distribution ; z is the input noise; is the prior distribution of the input noise, for the samples generated by the generator; represents is sampled from the prior distribution of the input noise ; is the weather prediction data generated by the generator according to the noise z; is the probability output by the discriminator, indicating the input data The probability that it is real weather data; is the output of the discriminator for the generated data indicating the probability that the generated data is judged as real weather data.

[0013] Furthermore, by using the adversarial weather prediction network to predict sudden or extreme weather patterns, the weather prediction results are combined with the power prediction for adversarial weather-power joint prediction.

[0014] An adversarial weather-power joint prediction model is established by processing multi-site collaborative prediction through online learning and transfer learning; the online learning updates the model parameters using the basic formula of stochastic gradient descent, expressed as: ; where is the model parameter at time point ; is the updated model parameter; is the learning rate at time point ; is the gradient of the loss function with respect to the parameter on the data ; is the new data collected at time point .

[0015] The is set as an adaptive learning rate, expressed as: ; where is the learning rate at time point ; is the initial learning rate; is the time step or the number of iterations; is the square root of the time step; by applying the adaptive learning rate, as the learning rate gradually decreases over time, the model learns quickly in the early stage and gradually stabilizes in the later stage.

[0016] The learning functional form of the transfer learning is expressed as: ; where is the maximum mean discrepancy loss; is the feature mapping function; is the th sample in the source domain; is the th sample in the target domain; is the number of samples in the source domain; is the number of samples in the target domain; is the norm in the reproducing kernel Hilbert space.

[0017] Based on the same inventive concept, on the other hand, the present invention also provides a new energy photovoltaic power prediction system based on deep learning. The system includes: a first data acquisition module, a second time series prediction module, a third weather prediction network module, and an adversarial weather-power joint prediction module, and the modules are connected in sequence.

[0018] The first data acquisition module is used to acquire historical power data and meteorological data of a photovoltaic site; preprocess the historical power data and meteorological data to obtain standardized data.

[0019] The second time series prediction module is used to extract local pattern features of meteorological data and power data based on the standardized data, and perform power prediction for the current time series through a single-site power prediction model; capture the spatial correlation between sites through a multi-site joint prediction model to predict the power of the photovoltaic site.

[0020] The third weather prediction network module is used to establish a weather pattern recognition method according to the meteorological data to distinguish normal weather patterns and sudden weather patterns; establish an adversarial weather prediction network based on adversarial learning technology.

[0021] The adversarial weather-power joint prediction module is used to establish an adversarial weather-power joint prediction model based on the adversarial weather prediction network. The adversarial weather-power joint prediction model performs multi-site power collaborative prediction through joint online learning and transfer learning of meteorological conditions and power output.

[0022] Advantageous Effects Compared with the prior art, the advantageous effects of the present invention are: 1. The prediction model can simultaneously process time series dependence and spatial correlation to improve the prediction ability.

[0023] 2. By learning the system characteristics in a data-driven manner, the adaptability is stronger and the prediction accuracy is further improved.

[0024] 3. The online learning and transfer learning mechanisms enable the system to continuously evolve and improve the stability of the long-term prediction accuracy. Description of the Drawings

[0025] Figure 1 It is a flow block diagram of a new energy photovoltaic power prediction method according to Embodiment 1 of the present invention; Figure 2 It is a schematic diagram of the module composition of a new energy photovoltaic power prediction system according to Embodiment 2 of the present invention. Detailed Embodiments

[0026] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0027] Embodiment 1: As Figure 1 shown, this embodiment provides a new energy photovoltaic power prediction method based on deep learning. The method includes the following steps: Step S1, obtaining historical power data and meteorological data of a photovoltaic site; the photovoltaic site is a single photovoltaic site or multiple photovoltaic sites within the research scope; the meteorological data includes temperature, humidity, wind speed, and light intensity; performing data preprocessing on the historical power data and meteorological data to obtain standardized data.

[0028] For example: The team collected historical data of three photovoltaic power stations in different geographical locations over the past two years. The data includes power generation data recorded every 15 minutes, as well as corresponding meteorological data (temperature, humidity, wind speed, and light intensity). One piece of data is as follows: The record of a certain station at a certain moment shows that the power is 8.2 megawatts, the temperature is 28 °C, the humidity is 45%, the wind speed is 3.5 m / s, and the light intensity is 950 W / m². After the data collection is completed, the team performs data preprocessing. Since the measurement ranges of different sensors are different, such as the temperature range is from -10 °C to 40 °C, while the power range is from 0 to 10 megawatts, the team performs data standardization processing. Standardize the temperature of 28 °C to 0.76 (calculated according to the minimum temperature of -10 °C and the maximum temperature of 40 °C), and standardize the power of 8.2 megawatts to 0.82 (calculated according to the minimum power of 0 megawatts and the maximum power of 10 megawatts).

[0029] Step S2, extracting local pattern features of meteorological data and power data based on the standardized data, and performing power prediction for the current time series through a single-site power prediction model; the single-site power prediction model includes a preset time series prediction model and a preset spatial feature extraction model; capturing the spatial correlation between sites through establishing a multi-site joint prediction model to predict the power of the photovoltaic site.

[0030] The preset spatial feature extraction model identifies and extracts local features of meteorological data and power data by learning the standardized data; the preset spatial feature extraction model uses a convolutional neural network for calculation, and the formula is: ; where is the value of the th feature map at the position ; is the value of the feature map of the layer at the position is the weight of the convolutional kernel of the layer at the position is the height and width of the convolutional kernel; is the configuration item of the is the preset activation function; M is the number of rows of the convolutional kernel, representing the number of pixels covered by the convolutional kernel in the vertical direction; N is the number of columns of the convolutional kernel, representing the number of pixels covered by the convolutional kernel in the horizontal direction.

[0031] It should be noted that local feature extraction is performed on meteorological and power data through a sliding window mechanism. The model adopts a multi-layer convolutional structure, and each layer of convolutional operation captures local spatial features and patterns in the data by combining weighted summation and a non-linear activation function. The convolutional kernel, as a feature detector, can identify feature patterns of different scales and types, such as temperature change trends, light intensity fluctuations, etc. The bias term further enhances the expression ability of the model and can adapt to data distributions with different baseline levels. The activation function introduces non-linearity, enabling the model to learn complex non-linear mapping relationships, thereby accurately extracting the complex correlation patterns between meteorological conditions and power output in the photovoltaic system.

[0032] By establishing the single-site power prediction model, the long-term and short-term temporal dependencies of the photovoltaic site are captured for power prediction at the current time series; first, the local feature values of meteorological data and power data are received as input features, and the historical learning information is used as the hidden state; subsequently, the forget gate is used to determine whether the hidden state information needs to be retained or discarded, and the input gate is used to filter the new information in the input features and generate a candidate memory unit; the memory unit is updated by weighted fusion of the forgotten hidden state information and the new information; finally, the power prediction value of the photovoltaic site at the current moment is generated based on the updated new memory unit.

[0033] For example: The system receives the input features at a certain time point: normalized temperature 0.74, humidity 0.46, wind speed 0.35, light intensity 0.92, and the historical learning information of the previous few time points. Through the forget gate mechanism, the system determines that the data weight at the same time point the previous day should be reduced (because the cloud cover is significantly different today), while retaining the data pattern in the morning today. Through the input gate, the system identifies the new information that the current light intensity is rising rapidly, predicts that the power at time point 1 will reach 8.2 MW, with a difference of only 0.1 MW from the actual value, and the prediction accuracy reaches 98.8%.

[0034] The method for establishing the multi-site joint prediction model includes: The multi-site joint prediction model updates the power information of the photovoltaic site to be predicted by studying the spatial correlation between multiple sites. The formula is as follows: ; where, is the feature representation of node at the th layer; represents the neighbor set of node ; represents the number of neighbors; represents the weight matrix of the th layer; is the configuration item of the th layer; is the feature representation of node at the th layer; is a preset activation function.

[0035] It should be noted that the multi-site joint prediction model of the present invention is based on the principle of graph neural network. Multiple photovoltaic sites are regarded as nodes in the graph structure, and the correlation between sites is mapped to the connection relationship between nodes. Through the aggregation and update mechanism of neighbor node information, the modeling of spatial correlation between sites is realized. The multi-site joint prediction effectively utilizes the similarity of meteorological conditions between sites with similar geographical locations, and significantly improves the prediction accuracy.

[0036] For example: When analyzing the previous data, it is found that the power of a certain station 1 drops at 9:00 am (from 5.6 MW to 4.2 MW). Through the correlation analysis between sites, the system identifies this as a signal of a moving cloud layer. Based on the geographical location relationship and historical weather patterns of the three stations, the system predicts that the cloud layer will affect a certain station 2 in about 45 minutes.

[0037] During the calculation process, the system assigns different weights to the three stations respectively: Since station 1 is the closest to station 2 (15 km), the highest weight of 0.6 is assigned; station 3 is farther away (25 km), and the weight of 0.4 is assigned. The system integrates this information and predicts that the power of station 2 at 9:45 will drop from the originally expected 7.8 MW to 6.3 MW. The actual record shows that the actual power at 9:45 is 6.5 MW, and the prediction error is only 3%.

[0038] Step S3, establish a weather pattern recognition method based on the meteorological data to distinguish normal weather patterns and sudden weather patterns; establish an adversarial weather prediction network based on adversarial learning technology, and the adversarial weather prediction network is a deep learning framework that generates weather prediction data through adversarial training.

[0039] The weather pattern recognition method classifies different weather patterns by constructing a hyperplane, sets a preset decision boundary, and updates the hyperplane through data iteration to update the preset decision boundary; the normal weather pattern and the sudden weather pattern are distinguished by the preset decision boundary after iteration.

[0040] The adversarial training is a mutual game process between the generator and the discriminator: the generator continuously improves the quality of the generated weather data, and the discriminator improves the ability to distinguish between real and generated data; the objective function of the adversarial weather prediction network is expressed as: ; where is the generator network; represents minimizing the generator error; is the discriminator network; represents maximizing the discriminator error; represents the objective functions of the generator and the discriminator; is a real data sample; is the distribution of real weather data; is the expected value; represents is sampled from the real weather data distribution ; z is the input noise; is the prior distribution of the input noise, for the samples generated by the generator; represents is sampled from the prior distribution of the input noise ; is the weather prediction data generated by the generator according to the noise z; is the probability output by the discriminator, indicating that the input data is the possibility of real weather data; is the output of the discriminator for the generated data indicating the probability that the generated data is judged as real weather data.

[0041] For example: On a certain afternoon, the system detected that the data reported by the weather station showed that there was a fast-moving cloud cluster in the northwest direction, and the wind speed increased from 2.3 m / s to 6.8 m / s in the past 20 minutes. The system immediately identified this as a sudden weather pattern. Subsequently, the adversarial weather prediction network started to work. Through the cooperation of the generator and the discriminator, this network generated high-precision weather prediction data for the next 3 hours. The generator predicted the moving path of the cloud cluster and the change in light intensity based on the current observations and historical similar weather patterns. The discriminator evaluated the credibility of the generated results by comparing with historical real data, prompting the generator to continuously improve the prediction quality. Finally, the system predicted that the cloud cluster would block the sun between 15:30 and 16:45, resulting in the light intensity dropping from 870 W / m² to 420 W / m², and further causing a significant drop in power output.

[0042] Step S4, based on the adversarial weather prediction network, establish an adversarial weather-power joint prediction model, which is a multi-site power collaborative prediction through joint online learning and transfer learning of meteorological conditions and power output.

[0043] Predict sudden or extreme weather patterns through the adversarial weather prediction network, and combine the weather prediction results with power prediction for adversarial weather-power joint prediction.

[0044] Establish an adversarial weather-power joint prediction model through online learning and transfer learning for multi-site collaborative prediction; the online learning updates the model parameters using the basic formula of stochastic gradient descent, expressed as: ; where is the model parameter at time point ; is the updated model parameter; is the learning rate at time point ; is the gradient of the loss function with respect to the parameter on the data ; is the new data collected at time point .

[0045] The is set as an adaptive learning rate, expressed as: ; where is the learning rate at time point ; is the initial learning rate; is the time step or number of iterations; is the square root of the time step; through the application of the adaptive learning rate, as the learning rate gradually decreases over time, the model learns quickly in the early stage and gradually stabilizes in the later stage.

[0046] It should be noted that the online learning mechanism of the present invention is based on the stochastic gradient descent algorithm, and adapts to the changing meteorological environment and the operating state of the photovoltaic system through real-time parameter updates. This online learning method is particularly suitable for application scenarios such as photovoltaic power prediction where the data distribution changes dynamically over time. The gradient calculation process takes into account the current parameter state and the characteristics of new data to ensure that the update direction is towards reducing the prediction error.

[0047] The learning functional expression of the transfer learning is represented as: ; Wherein, is the maximum mean discrepancy loss; is the feature mapping function; is the th sample of the source domain; is the th sample of the target domain; is the number of source domain samples; is the number of target domain samples; is the norm in the reproducing kernel Hilbert space.

[0048] It should be noted that the transfer learning mechanism realizes the effective transfer of knowledge between different photovoltaic sites through the maximum mean discrepancy loss function. The feature mapping function maps the source domain and target domain data into a high-dimensional feature space, and the maximum mean discrepancy measures the difference in distribution between the two domains in this space. By minimizing this difference, the model can effectively transfer the knowledge learned at data-rich sites to data-sparse sites, solving the problem of insufficient data at newly built or small photovoltaic sites; the norm in the reproducing kernel Hilbert space provides a strict mathematical measure of the distribution difference. The transfer learning technology greatly reduces the dependence on the historical data of the target site, enabling the model to still maintain high-precision prediction ability under limited data conditions, and is applicable to the actual application scenarios with diverse scales and wide distributions of photovoltaic sites.

[0049] Embodiment 2: Based on the same inventive concept, as Figure 2 shown, this embodiment also provides a new energy photovoltaic power prediction system based on deep learning, and the system includes: a first data acquisition module, a second time series prediction module, a third weather prediction network module, and an adversarial weather-power joint prediction module, and the modules are connected in sequence.

[0050] The first data acquisition module is used to acquire the historical power data and meteorological data of the photovoltaic site; perform data preprocessing on the historical power data and meteorological data to obtain standardized data.

[0051] The second time series prediction module is used to extract local pattern features of meteorological data and power data based on the standardized data, and perform power prediction for the current time series through a single-site power prediction model; a multi-site joint prediction model is established to capture the spatial correlation between sites to predict the power of photovoltaic sites.

[0052] The third weather prediction network module is used to establish a weather pattern recognition method based on the meteorological data to distinguish normal weather patterns and sudden weather patterns; an adversarial weather prediction network is established based on adversarial learning technology.

[0053] The adversarial weather-power joint prediction module is used to establish an adversarial weather-power joint prediction model based on the adversarial weather prediction network. The adversarial weather-power joint prediction model performs multi-site power collaborative prediction through joint online learning and transfer learning of meteorological conditions and power output.

[0054] It should be noted that regarding the system in the above embodiments, the specific manners in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0055] Finally, it should be noted that although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A new energy photovoltaic power prediction method based on deep learning, characterized in that, The method includes the following steps: Step S1, obtaining historical power data and meteorological data of a photovoltaic site; the photovoltaic site is a single photovoltaic site or multiple photovoltaic sites within the research scope; the meteorological data includes temperature, humidity, wind speed, and light intensity; performing data preprocessing on the historical power data and meteorological data to obtain standardized data; Step S2, extracting local pattern features of meteorological data and power data based on the standardized data, and performing power prediction for the current time series through a single-site power prediction model; the single-site power prediction model includes a preset time series prediction model and a preset spatial feature extraction model; capturing the spatial correlation between sites through establishing a multi-site joint prediction model to predict the power of the photovoltaic site; Step S3, establishing a weather pattern recognition method based on the meteorological data to distinguish normal weather patterns and sudden weather patterns; establishing an adversarial weather prediction network based on adversarial learning technology, and the adversarial weather prediction network is a deep learning framework that generates weather prediction data through adversarial training; Step S4, based on the adversarial weather prediction network, establishing an adversarial weather-power joint prediction model, and the adversarial weather-power joint prediction model performs multi-site power collaborative prediction through joint online learning and transfer learning of meteorological conditions and power output.

2. The new energy photovoltaic power prediction method based on deep learning according to claim 1, wherein The preset spatial feature extraction model identifies and extracts local features of meteorological data and power data by learning the standardized data; the preset spatial feature extraction model uses a convolutional neural network for calculation, and the formula is: ； Among them, is the value of the -th layer feature map at position ; is the value of the -th layer feature map at position ; is the weight of the -th layer convolution kernel at position ; is the height and width of the convolution kernel; is the configuration item of the -th layer; is the preset activation function; M is the number of rows of the convolution kernel, representing the number of pixels covered by the convolution kernel in the vertical direction; N is the number of columns of the convolution kernel, representing the number of pixels covered by the convolution kernel in the horizontal direction.

3. A new energy photovoltaic power prediction method based on deep learning according to claim 1, characterized in that, The method for power prediction includes: Capturing the long-term and short-term time series dependencies of the photovoltaic site through establishing the single-site power prediction model to perform power prediction for the current time series; first, receiving the local feature values of meteorological data and power data as input features, and using the historical learning information as the hidden state; subsequently, judging whether the hidden state information needs to be retained or discarded through the forgetting gate, screening the new information in the input features through the input gate and generating a candidate memory unit; updating the memory unit by weighted fusion of the forgotten hidden state information and the new information; finally, generating the power prediction value of the photovoltaic site at the current moment according to the updated new memory unit.

4. A new energy photovoltaic power prediction method based on deep learning according to claim 3, characterized in that The method for establishing the multi-site joint prediction model includes: The multi-site joint prediction model updates the power information of the photovoltaic site to be predicted by studying the spatial correlation between multiple sites, and the formula is: ； Among them, is the feature representation of the node at the layer; represents the neighbor set of the node ; represents the number of neighbors; represents the weight matrix of the layer; is the configuration item of the layer; is the feature representation of the node at the layer; is a preset activation function.

5. A new energy photovoltaic power prediction method based on deep learning according to claim 1, characterized in that The weather pattern recognition method classifies different weather patterns by constructing a hyperplane, sets a preset decision boundary, and updates the hyperplane and then the preset decision boundary through data iteration; distinguishing normal weather patterns and sudden weather patterns through the preset decision boundary after iteration; The adversarial training is a mutual game process between a generator and a discriminator: the generator continuously improves the quality of the generated weather data, and the discriminator improves the ability to distinguish real and generated data; the objective function of the adversarial weather prediction network is expressed as: ； Among them, is the generator network; represents minimizing the generator error; is the discriminator network; represents maximizing the discriminator error; represents the objective functions of the generator and the discriminator; is the real data sample; is the distribution of real weather data; is the expected value; represents is sampled from the real weather data distribution ; z is the input noise; is the prior distribution of the input noise for the samples generated by the generator; represents that z is sampled from the prior distribution of the input noise ; is the weather prediction data generated by the generator according to the noise z; is the probability output by the discriminator, indicating that the input data is the possibility of real weather data; is the output of the discriminator for the generated data indicating the probability that the generated data is judged as real weather data.

6. The method for predicting new energy photovoltaic power based on deep learning according to claim 1, wherein, Performing sudden or extreme weather pattern prediction through the adversarial weather prediction network, and combining the weather prediction result with the power prediction for adversarial weather-power joint prediction; Establish an adversarial weather-power joint prediction model by processing multi-site collaborative prediction through online learning and transfer learning; the online learning updates the model parameters using the basic formula of stochastic gradient descent, expressed as: ; where is the model parameter at time point ; is the updated model parameter; is the learning rate at time point ; is the gradient of the loss function with respect to the parameter on the data ; is the new data collected at time point ; The said is set to an adaptive learning rate, expressed as: ; where is the learning rate at time point ; is the initial learning rate; is the time step or the number of iterations; is the square root of the time step; By applying the said adaptive learning rate, as the learning rate gradually decreases over time, the model learns quickly in the early stage and gradually stabilizes in the later stage; The learning function of the transfer learning is represented as follows: ； Among them, is the maximum mean discrepancy loss; is the feature mapping function; is the -th sample in the source domain; is the -th sample in the target domain; is the number of source domain samples; is the number of target domain samples; is the norm in the reproducing kernel Hilbert space.

7. A new energy photovoltaic power prediction system based on deep learning, for performing the method according to any one of claims 1-6, characterized in that, The system includes: a first data acquisition module, a second time series prediction module, a third weather prediction network module, and an adversarial weather-power joint prediction module, and the modules are connected in sequence; The first data acquisition module is used to acquire the historical power data and meteorological data of the photovoltaic site; preprocess the historical power data and meteorological data to obtain standardized data; The second time series prediction module is used to extract the local pattern features of the meteorological data and power data based on the standardized data, and perform power prediction for the current time series through a single-site power prediction model; capture the spatial correlation between sites through a multi-site joint prediction model to predict the power of the photovoltaic site; The third weather prediction network module is used to establish a weather pattern recognition method according to the meteorological data to distinguish between normal weather patterns and sudden weather patterns; establish an adversarial weather prediction network based on adversarial learning technology; The adversarial weather-power joint prediction module is used to establish an adversarial weather-power joint prediction model based on the adversarial weather prediction network. The adversarial weather-power joint prediction model performs multi-site power collaborative prediction through joint online learning and transfer learning of meteorological conditions and power output.

Images