Multi-source network type microgrid photovoltaic power generation prediction method based on deep learning

Abstract

The application relates to the technical field of deep learning, and particularly discloses a deep learning-based multi-source network type micro-grid photovoltaic power generation capacity prediction method, which comprises the following steps: obtaining historical power generation capacity data and historical weather data of a photovoltaic power generation system in a micro-grid, and preprocessing the historical power generation capacity data and the historical weather data; calculating a photovoltaic power generation potential value based on the preprocessed historical weather data; taking the preprocessed historical power generation capacity data, the historical weather data and the photovoltaic power generation potential value as input features to construct a training data set, and training a deep learning model by using the training data set; and using the trained deep learning model to predict photovoltaic power generation capacity and output. The application can output future power generation capacity which is more in line with real power generation scale, is favorable for improving the decision accuracy of the micro-grid in power dispatching and energy management, and enhances the stability and efficiency of system operation.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power generation prediction technology, specifically to a method for predicting photovoltaic power generation in multi-source grid-type microgrids based on deep learning. Background Technology

[0002] In microgrid environments, accurate forecasting of photovoltaic power generation is crucial for power dispatch and energy management. However, existing forecasting methods often rely on simple historical data or basic weather parameters, failing to fully consider the dynamic and complex nature of weather conditions, such as the interaction of factors like changes in solar position, temperature fluctuations, and cloud cover.

[0003] This leads to a significant discrepancy between the forecast results and actual power generation, making it impossible to effectively capture real-time changes in power generation potential. Since microgrids require high-precision power generation forecasts to optimize power distribution and avoid supply-demand imbalances, this inaccuracy directly affects system stability and efficiency, increases operational risks, and limits the efficient integration of renewable energy. Summary of the Invention

[0004] The purpose of this application is to provide a method for predicting photovoltaic power generation in multi-source grid-type microgrids based on deep learning, thereby solving the aforementioned technical problems.

[0005] The objective of this application can be achieved through the following technical solutions: A deep learning-based method for predicting photovoltaic power generation in multi-source grid-connected microgrids includes the following steps: Acquire historical power generation data and historical weather data of photovoltaic power generation systems in microgrids, and preprocess the historical power generation data and historical weather data; Calculate the photovoltaic power generation potential value based on preprocessed historical weather data; A training dataset is constructed using preprocessed historical power generation data, historical weather data, and photovoltaic power generation potential values ​​as input features, and a deep learning model is trained using the training dataset. Using the trained deep learning model, predict and output the photovoltaic power generation.

[0006] As a further aspect of this application: obtaining historical power generation data and historical weather data, including: The historical power generation data is collected from the microgrid monitoring system. The historical power generation data is the actual output power data of the photovoltaic power generation system recorded in time series form. The historical weather data is obtained from the meteorological station. The historical weather data includes solar irradiance, temperature and cloud cover, where solar irradiance is global horizontal irradiance, temperature is ambient temperature and cloud cover is the percentage of cloud cover.

[0007] As a further aspect of this application, preprocessing of historical power generation data and historical weather data includes: Data cleaning is performed on historical power generation data. Data cleaning includes identifying outliers in the historical power generation data based on box plot methods and removing the identified outliers. The historical power generation data after data cleaning is normalized using the min-max normalization method to normalize the historical power generation data to the range of zero to one. Repeat the above process to complete the preprocessing of historical weather data.

[0008] As a further aspect of this application: calculating the photovoltaic power generation potential value includes: Based on the geographical location, date, and time of the microgrid, the solar altitude angle and azimuth angle are calculated. Based on the solar altitude angle and azimuth angle, the theoretical maximum solar irradiance is calculated using a solar irradiance model. The ratio of the preprocessed solar irradiance to the theoretical maximum solar irradiance is calculated and used as the irradiance ratio. Calculate the difference between the pre-processed temperature and a preset standard temperature, and multiply the difference by a preset coefficient to obtain the temperature influence factor; The photovoltaic power generation potential value is obtained by multiplying the irradiance ratio by the temperature influence factor.

[0009] As a further aspect of this application: constructing a training dataset, including: The preprocessed historical power generation data, historical weather data, and photovoltaic power generation potential value are aligned in chronological order to obtain time-aligned data; each time point in the time-aligned data contains input features, which include the preprocessed historical power generation data, the preprocessed historical weather data, and the photovoltaic power generation potential value; Based on a preset input window length, the time-aligned data is divided into multiple consecutive input time windows using a sliding window method, with each input time window containing input features from multiple consecutive time points. For each input time window, based on the preset prediction target time length, preprocessed historical power generation data at multiple time points after the end time of the input time window are obtained from the time-aligned data as prediction targets. The input features of each input time window are combined with the corresponding prediction target to form a training sample; Combine all training samples to obtain the training dataset.

[0010] As a further aspect of this application: training a deep learning model includes: The input features from the training dataset are input into the input layer of the deep learning model, which includes an input layer, multiple hidden layers, and an output layer, wherein the hidden layers adopt a long short-term memory network layer structure. The input features are processed through the multiple hidden layers, and the temporal dependencies in the input features are captured based on the long short-term memory network structure to obtain the hidden layer output. The output layer processes the output of the hidden layer to generate multiple power generation prediction values. These multiple power generation prediction values ​​correspond to the time points of the prediction target in the training dataset, and each power generation prediction value represents the predicted power generation at a specific future time point. Based on the multiple predicted power generation values ​​and the prediction target in the training dataset, the prediction error is calculated using the mean squared error loss function. The calculation of the prediction error includes taking the squared difference between the multiple predicted power generation values ​​and the values ​​at corresponding time points in the prediction target, and calculating the average of all squared differences to obtain the prediction error. Based on the prediction error, the gradient of the parameters of each layer in the deep learning model is calculated using the backpropagation algorithm; based on the gradient, the parameters of the deep learning model are updated using an optimizer to obtain the updated deep learning model. Repeat the above training steps until the preset training stopping condition is met to obtain the trained deep learning model.

[0011] As a further aspect of this application: predicting future photovoltaic power generation, including: Starting from the current moment, a window with a length equal to the length of the input window is set along the counter-time axis, denoted as the target window. The input features within the target window are used as the real-time input features. The real-time input features are input into the trained deep learning model, and the real-time input features are processed through multiple hidden layers of the deep learning model to obtain the real-time hidden layer output. The output of the real-time hidden layer is processed by the output layer of the deep learning model to generate multiple power generation prediction values, which correspond to multiple time points within a preset prediction time length. The difference A between the historical maximum and historical minimum values ​​of the historical power generation data is obtained. The predicted power generation value is multiplied by the difference A and then added to the historical minimum value to obtain the photovoltaic power generation at a future point in time.

[0012] Implementing one of the technical solutions described in this application has the following advantages or beneficial effects: This application, by constructing a comprehensive input feature system that includes historical power generation data, weather information, and photovoltaic power generation potential values ​​obtained based on solar irradiance and temperature characteristics, can more comprehensively reflect the dynamic changes in the photovoltaic power generation process and achieve a refined characterization of power generation behavior. By cleaning, normalizing, and time-aligning the data, and using a sliding window approach to construct training samples, the model can accurately learn the trends and short-term fluctuations in the time series. Simultaneously, by utilizing the temporal structure in the deep learning model to capture the temporal dependencies between input features, the prediction results are made closer to actual power generation changes. The photovoltaic power generation potential values ​​generated by this application effectively enhance the model's sensitivity to changes in weather conditions, enabling the prediction to reflect the combined effects of irradiance and temperature, thereby improving the prediction results' responsiveness to actual photovoltaic power generation trends. Through inverse normalization of the prediction results, this application can output future power generation that more closely matches the actual power generation scale, which is beneficial for improving the decision-making accuracy of microgrids in power dispatch and energy management, and enhancing the stability and efficiency of system operation. Attached Figure Description

[0013] The present application will be further described below with reference to the accompanying drawings.

[0014] Figure 1 This is a flowchart illustrating the photovoltaic power generation prediction method for multi-source grid-type microgrids based on deep learning, as proposed in this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, various exemplary embodiments described below will be referenced to the accompanying drawings, which form part of the exemplary embodiments and depict various exemplary embodiments that may be adopted to implement this application. Unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. It should be understood that they are merely examples of processes, methods, and apparatuses consistent with some aspects of this application disclosed as detailed in the appended claims, and other embodiments may be used, or structural and functional modifications may be made to the embodiments listed herein without departing from the scope and spirit of this application.

[0016] In the description of this application, it should be understood that the terms "center," "longitudinal," "lateral," etc., indicate the orientation or positional relationship based on the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the referred element must have a specific orientation, or be constructed and operated in a specific orientation. The terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. The term "multiple" means two or more. The terms "connected" and "linked" should be interpreted broadly, for example, they can be fixed connections, detachable connections, integral connections, mechanical connections, electrical connections, communication connections, direct connections, indirect connections through an intermediate medium, and can be the internal connection of two elements or the interaction relationship between two elements. The term "and / or" includes any and all combinations of one or more of the related listed items. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0017] Please see Figure 1 As shown, this application presents a method for predicting photovoltaic power generation in multi-source grid-connected microgrids based on deep learning, comprising the following steps: Acquire historical power generation data and historical weather data of photovoltaic power generation systems in microgrids, and preprocess the historical power generation data and historical weather data; A preferred embodiment of this application involves acquiring historical power generation data and historical weather data, including: The historical power generation data is collected from the microgrid monitoring system. The historical power generation data is the actual output power data of the photovoltaic power generation system recorded in time series form. The historical weather data is obtained from the meteorological station. The historical weather data includes solar irradiance, temperature and cloud cover, where solar irradiance is global horizontal irradiance, temperature is ambient temperature and cloud cover is the percentage of cloud cover.

[0018] Another preferred embodiment of this application involves preprocessing historical power generation data and historical weather data, including: Data cleaning is performed on historical power generation data. Data cleaning includes identifying outliers in the historical power generation data based on box plot methods and removing the identified outliers. The historical power generation data after data cleaning is normalized using the min-max normalization method to normalize the historical power generation data to the range of zero to one. Repeat the above process to complete the preprocessing of historical weather data.

[0019] It should be noted that the preprocessing process for historical weather data is analogous to that for historical power generation data, as both use the exact same processing flow, and will not be elaborated upon here.

[0020] Calculate the photovoltaic power generation potential value based on preprocessed historical weather data; In another preferred embodiment of this application, calculating the photovoltaic power generation potential value includes: The calculation of photovoltaic power generation potential is based on time and geographical information. By obtaining the latitude and longitude of the microgrid's location, along with the corresponding date and time, the spatial position of the sun at that moment is derived from this input data. The solar altitude angle describes the sun's vertical position relative to the horizon, reflecting the accessibility of solar energy, while the azimuth angle reflects the sun's projected position in the horizontal direction. These two angles are derived based on the astronomical description of the apparent motion of the sun caused by the Earth's rotation and revolution, through continuous transformations of geographical latitude, solar declination angle, and hour angle. The larger the solar altitude angle, the more theoretically available irradiance energy can be obtained. Therefore, after obtaining the solar altitude angle and azimuth angle, they are used as input variables in the solar irradiance model. Through the model's internal processing of the sun's direct path, atmospheric attenuation characteristics, and the relationship of solar angular projection, the theoretical maximum solar irradiance under cloudless and standard atmospheric conditions can be obtained.

[0021] After obtaining the theoretical maximum irradiance, the preprocessed solar irradiance data is input and compared with the theoretical value to obtain the proportional relationship between the two. The higher the ratio, the closer the actual irradiance is to the ideal state. Then, the preprocessed temperature data is input, and the difference is calculated with the set standard temperature. This difference is then multiplied by a preset coefficient to reflect the impact of temperature changes on power generation potential in a linear weighted manner.

[0022] Temperature affects the performance of photovoltaic modules, so this factor is used to describe the correction of the potential value when the temperature deviates from the optimal operating conditions. After calculating the irradiance ratio and the temperature influence factor, the two are multiplied to form the final photovoltaic power generation potential value. The output is a numerical feature used for subsequent model training, reflecting the power generation capacity level that the photovoltaic system may achieve at a specific time.

[0023] It is important to note that the construction of photovoltaic (PV) power generation potential relies on the combination of solar geometry and environmental characteristics. The principle is that the solar altitude angle and azimuth angle jointly determine the path direction of light reaching the PV module, thus affecting the available effective light energy. The theoretical maximum irradiance calculated using this relationship provides an idealized reference, making the subsequent understanding of actual irradiance conditions more physically meaningful. The ratio between the actual irradiance and this theoretical value reflects the degree of light energy reduction caused by weather conditions, quantifying the impact of cloud cover and atmospheric changes on sunlight. The introduction of a temperature influence factor is based on the variation of the electrical performance of PV materials at different temperatures. By combining the degree of temperature deviation from the optimal operating point with the influence coefficient, the performance changes of the PV module are expressed in a measurable way. Multiplying the irradiance ratio and the temperature influence factor aims to utilize the principle that the combined effect of light and temperature determines PV power generation capacity, ensuring that the output PV power generation potential value simultaneously reflects the sufficiency of light conditions and the performance deviation caused by temperature, thus more closely reflecting the actual mechanism of environmental influence on PV power generation.

[0024] A training dataset is constructed using preprocessed historical power generation data, historical weather data, and photovoltaic power generation potential values ​​as input features, and a deep learning model is trained using the training dataset. In a preferred embodiment of this application, a training dataset is constructed, comprising: The preprocessed historical power generation data, historical weather data, and photovoltaic power generation potential value are aligned in chronological order to obtain time-aligned data; each time point in the time-aligned data contains input features, which include the preprocessed historical power generation data, the preprocessed historical weather data, and the photovoltaic power generation potential value; Based on a preset input window length, the time-aligned data is divided into multiple consecutive input time windows using a sliding window method, with each input time window containing input features from multiple consecutive time points. For each input time window, based on the preset prediction target time length, preprocessed historical power generation data at multiple time points after the end time of the input time window are obtained from the time-aligned data as prediction targets. The input features of each input time window are combined with the corresponding prediction target to form a training sample; Combine all training samples to obtain the training dataset.

[0025] It is understandable that after windowing, the multi-source feature sequences aligned by time can enable deep learning models to learn the temporal correlation structure with a fixed-length continuous input observation range. The principle is that photovoltaic power generation is affected by weather and environmental factors and exhibits obvious temporal dependence characteristics. There is a causal relationship between power generation and weather changes in continuous time periods, and the model needs to capture this correlation pattern that evolves over time in data fragments of a uniform format.

[0026] By dividing a long-term series into multiple independent continuous segments through a sliding window, the model obtains a consistent input structure in each segment. This allows the model to repeatedly see patterns with the same structure in different time periods during training, thereby extracting universal features from the time patterns.

[0027] Several future time points after the end of the window are used as prediction targets because time series prediction relies on the indicative role of historical periods in the future state. The window provides historical conditions, and the future power generation as a target reflects the influence of these conditions on future trends, enabling the training samples to have a learnable mapping relationship from input to output.

[0028] By combining each window and its corresponding prediction target into training samples, the model can essentially learn how continuous historical features affect future power generation performance, forming a time-dependent learning framework. The training dataset consists of a large number of samples with similar structures, enabling the model to extract stable temporal patterns through repeated learning and strengthen its understanding of photovoltaic power generation trends, thus providing a reliable foundation for subsequent prediction tasks.

[0029] Another preferred embodiment of this application, training a deep learning model, includes: The input features from the training dataset are input into the input layer of the deep learning model, which includes an input layer, multiple hidden layers, and an output layer, wherein the hidden layers adopt a long short-term memory network layer structure. The input features are processed through the multiple hidden layers, and the temporal dependencies in the input features are captured based on the long short-term memory network structure to obtain the hidden layer output. The output layer processes the output of the hidden layer to generate multiple power generation prediction values. These multiple power generation prediction values ​​correspond to the time points of the prediction target in the training dataset, and each power generation prediction value represents the predicted power generation at a specific future time point. Based on the multiple predicted power generation values ​​and the prediction target in the training dataset, the prediction error is calculated using the mean squared error loss function. The calculation of the prediction error includes taking the squared difference between the multiple predicted power generation values ​​and the values ​​at corresponding time points in the prediction target, and calculating the average of all squared differences to obtain the prediction error. Based on the prediction error, the gradient of the parameters of each layer in the deep learning model is calculated using the backpropagation algorithm; based on the gradient, the parameters of the deep learning model are updated using an optimizer to obtain the updated deep learning model. Repeat the above training steps until the preset training stopping condition is met to obtain the trained deep learning model.

[0030] It is worth noting that the Long Short-Term Memory (LSTM) network structure can retain both short-term changes and trend information over a longer time span while processing input features. This ability stems from its internal state update method, which can transmit key information in the time dimension, enabling the model to effectively express historical change patterns when facing the continuous changes in photovoltaic power generation over time.

[0031] After the input layer receives the constructed data sequence, the hidden layer analyzes the sequential relationship between the input features based on the state update mechanism, enabling the model to deduce the possible direction of change in the future by utilizing the dependencies between each time point. Therefore, the information carried by the output of the hidden layer has the basis for inferring the future trend of power generation.

[0032] When processing the output of the hidden layer, the output layer establishes a mapping between the input sequence and the power generation at multiple future moments, enabling the model to learn how to transform historical features into future predictions during training.

[0033] The difference between the predicted value and the prediction target is measured by the mean squared error loss function, so that the network knows the source of the deviation in each iteration. The backpropagation algorithm calculates the influence of parameters on the error layer by layer along the network structure according to the deviation, so that the parameters are continuously adjusted in the direction of reducing the error. The network gradually forms a stable temporal mapping capability by updating the parameters through the optimizer. As the iteration proceeds, the model gradually converges in repeated error feedback and parameter correction, so that the deep learning model obtains a stable structure suitable for photovoltaic power generation prediction.

[0034] The entire process relies on time-series feature learning, error-driven weight adjustment, and iterative optimization mechanisms to enable the model to accurately grasp the relationship between the past and the future in the time series, thereby establishing a reliable basic representation capability for the prediction of photovoltaic power generation.

[0035] Using the trained deep learning model, predict and output the photovoltaic power generation.

[0036] Another preferred embodiment of this application predicts future photovoltaic power generation, including: Starting from the current moment, a window with a length equal to the length of the input window is set along the counter-time axis, denoted as the target window. The input features within the target window are used as the real-time input features. The real-time input features are input into the trained deep learning model, and the real-time input features are processed through multiple hidden layers of the deep learning model to obtain the real-time hidden layer output. The output of the real-time hidden layer is processed by the output layer of the deep learning model to generate multiple power generation prediction values, which correspond to multiple time points within a preset prediction time length. The difference A between the historical maximum and historical minimum values ​​of the historical power generation data is obtained. The predicted power generation value is multiplied by the difference A and then added to the historical minimum value to obtain the photovoltaic power generation at a future point in time.

[0037] It is important to note that constructing the target window by looking back from the current moment is a requirement arising from the dependence of time series prediction on continuous historical information. The arrangement of historical data in time sequence determines the basis for the formation of future states. Therefore, the target window can provide a real-time input with a complete temporal structure to the trained deep learning model, enabling the model to generate inferences about future trends based on continuous features.

[0038] After real-time input features are fed into the deep learning model, the hidden layer integrates historical trend information with short-term fluctuations by processing the temporal features. This enables the model to reconstruct the internal representation of the current state at the prediction time. This representation inherits the understanding of temporal patterns formed by the model during the training phase, and thus can play a role in the prediction phase.

[0039] After receiving the real-time output from the hidden layer, the output layer generates predicted values ​​for multiple future time points, allowing the power generation trends at different future time locations to be expressed simultaneously in a single inference. The predicted values ​​are the results under normalized conditions, so it is necessary to combine the difference between the historical maximum and minimum values ​​to map them back to the original dimensions. Through this inverse normalization method, the predicted values ​​are restored to a scale consistent with the actual power generation, enabling the model output to be directly used in practical applications.

[0040] The entire inference process utilizes historical windows, time-series mapping capabilities formed during training, and inverse normalization processing to jointly support the generation of future power generation values. This enables the model to automatically provide future power generation change trends based on current inputs during the prediction phase, thereby ensuring that the prediction results are consistent with actual dimensions and achieving power generation output oriented towards actual dispatch needs.

[0041] This application introduces a photovoltaic (PV) power generation potential value—a feature that comprehensively reflects the combined effects of solar irradiance and temperature changes—into the deep learning prediction process. This allows the model to learn the patterns of power generation changes without relying solely on raw weather data. Instead, it enables the model to understand the upper limit of PV system power generation at a specific time based on an environmental capability index derived from physical laws. This feature enhances the model's responsiveness to dynamic weather changes, enabling it to more accurately depict the changing trends of PV power generation behavior under the combined influence of factors such as cloud cover variations, temperature fluctuations, and solar position shifts. Consequently, the prediction results exhibit higher sensitivity and adaptability. Traditional methods often cannot directly determine the power generation potential of a PV system from weather parameters under given conditions. However, this approach, by constructing a potential value, provides the model with auxiliary information reflecting the environment's power generation capacity. This allows the prediction to more closely reflect the actual physical changes in PV power generation. It is this combination of physically derived information and deep learning mechanisms that gives this application a unique effect in prediction precision, distinct from existing methods.

[0042] The above description is merely a preferred embodiment of this application. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of this application. Furthermore, under the teachings of this application, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of this application. Therefore, this application is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of this application.

Claims

1. A method for predicting photovoltaic power generation in multi-source grid-connected microgrids based on deep learning, characterized in that, Includes the following steps: Historical power generation data and historical weather data of the photovoltaic power generation system in the microgrid are obtained, and the historical power generation data and historical weather data are preprocessed. Calculate the photovoltaic power generation potential value based on preprocessed historical weather data; A training dataset is constructed using preprocessed historical power generation data, historical weather data, and photovoltaic power generation potential values ​​as input features, and a deep learning model is trained using the training dataset. Using the trained deep learning model, predict and output the photovoltaic power generation.

2. The photovoltaic power generation prediction method for multi-source grid-type microgrids based on deep learning according to claim 1, characterized in that, Obtain historical power generation data and historical weather data, including: The historical power generation data is collected from the microgrid monitoring system. The historical power generation data is the actual output power data of the photovoltaic power generation system recorded in time series form. The historical weather data is obtained from the meteorological station. The historical weather data includes solar irradiance, temperature, and cloud cover, wherein the solar irradiance is global horizontal irradiance, the temperature is ambient temperature, and the cloud cover is the percentage of cloud cover.

3. The photovoltaic power generation prediction method for multi-source grid-type microgrids based on deep learning according to claim 2, characterized in that, The historical power generation data and historical weather data are preprocessed, including: The historical power generation data is cleaned, including identifying outliers in the historical power generation data based on the box plot method and removing the identified outliers. The historical power generation data after data cleaning is normalized using the min-max normalization method to normalize the historical power generation data to the range of zero to one. Repeat the above process to complete the preprocessing of the historical weather data.

4. The photovoltaic power generation prediction method for multi-source grid-type microgrids based on deep learning according to claim 3, characterized in that, Calculating the potential value of photovoltaic power generation includes: Based on the geographical location, date, and time of the microgrid, the solar altitude angle and azimuth angle are calculated. Based on the solar altitude angle and azimuth angle, the theoretical maximum solar irradiance is calculated using a solar irradiance model. The ratio of the preprocessed solar irradiance to the theoretical maximum solar irradiance is calculated and used as the irradiance ratio. Calculate the difference between the pre-processed temperature and a preset standard temperature, and multiply the difference by a preset coefficient to obtain the temperature influence factor; The photovoltaic power generation potential value is obtained by multiplying the irradiance ratio by the temperature influence factor.

5. The photovoltaic power generation prediction method for multi-source grid-type microgrids based on deep learning according to claim 4, characterized in that, The construction of the training dataset includes: The preprocessed historical power generation data, historical weather data, and photovoltaic power generation potential value are aligned in chronological order to obtain time-aligned data; each time point in the time-aligned data contains input features, which include the preprocessed historical power generation data, the preprocessed historical weather data, and the photovoltaic power generation potential value; Based on a preset input window length, the time-aligned data is divided into multiple consecutive input time windows using a sliding window method, with each input time window containing input features from multiple consecutive time points. For each input time window, based on a preset prediction target time length, preprocessed historical power generation data at multiple time points after the end time of the input time window are obtained from the time-aligned data as prediction targets. The input features of each input time window are combined with the corresponding prediction target to form a training sample; Combine all training samples to obtain the training dataset.

6. The photovoltaic power generation prediction method for multi-source grid-type microgrids based on deep learning according to claim 5, characterized in that, Training the deep learning model includes: The input features from the training dataset are input into the input layer of the deep learning model, which includes an input layer, multiple hidden layers, and an output layer, wherein the hidden layers adopt a long short-term memory network layer structure. The input features are processed by multiple hidden layers, and the temporal dependencies in the input features are captured based on the long short-term memory network structure to obtain the hidden layer output. The output layer processes the output of the hidden layer to generate multiple power generation prediction values. These multiple power generation prediction values ​​correspond to the time points of the prediction target in the training dataset, and each power generation prediction value represents the predicted power generation at a specific future time point. Based on multiple predicted power generation values ​​and the prediction target in the training dataset, the prediction error is calculated using the mean squared error loss function. The calculation of the prediction error includes taking the squared difference between the multiple predicted power generation values ​​and the corresponding time point values ​​in the prediction target, and calculating the average of all squared differences to obtain the prediction error. Based on the prediction error, the gradient of the parameters of each layer in the deep learning model is calculated using the backpropagation algorithm; based on the gradient, the parameters of the deep learning model are updated using an optimizer to obtain the updated deep learning model. Repeat the above training steps until the preset training stopping condition is met to obtain the trained deep learning model.

7. The photovoltaic power generation prediction method for multi-source grid-type microgrids based on deep learning according to claim 6, characterized in that, Predicting future photovoltaic power generation, including: Starting from the current moment, a window with a length equal to the length of the input window is set along the counter-time axis, denoted as the target window. The input features within the target window are used as the real-time input features. The real-time input features are input into the trained deep learning model, and the real-time input features are processed through multiple hidden layers of the deep learning model to obtain the real-time hidden layer output. The output of the real-time hidden layer is processed by the output layer of the deep learning model to generate multiple power generation prediction values, which correspond to multiple time points within a preset prediction time length. The difference A between the historical maximum and historical minimum values ​​of the historical power generation data is obtained. The predicted power generation value is multiplied by the difference A and then added to the historical minimum value to obtain the photovoltaic power generation at a future point in time.

Images