Distributed photovoltaic power prediction method and system based on machine learning

Abstract

The invention relates to the technical field of photovoltaic power prediction, and provides a distributed photovoltaic power prediction method and system based on machine learning. The method comprises the following steps: interacting grid-connected position nodes and installation parameters of photovoltaic base stations in a photovoltaic jurisdiction area, and determining a distributed topology; mining photovoltaic power elements, taking modal decomposition-reconstruction-prediction as underlying logic, developing a power prediction module, and carrying out convergence learning and embedded connection; accessing a front-end acquisition card, and constraining and acquiring a sampling matrix of photovoltaic power elements through a timestamp; and importing the sampling matrix into a power prediction module, performing interval prediction decision through dynamic and static combination, and determining a power triple. According to the method, the technical problem of insufficient prediction precision caused by complex topological structure and data acquisition asynchronism of a distributed photovoltaic system is solved, and the technical effect of high-precision and robust power prediction is achieved through distributed topology construction, a modal decomposition-reconstruction prediction mechanism and dynamic and static combined interval prediction decision.

Description

Technical Field

[0001] The present application relates to the technical field of photovoltaic power prediction, and specifically to a distributed photovoltaic power prediction method and system based on machine learning. Background Art

[0002] With the acceleration of energy structure transformation, distributed photovoltaic systems have become an important component of the new power system due to their flexible access characteristics. However, the characteristics of distributed photovoltaic nodes, such as complex topology, strong asynchronous data collection, and significant power fluctuations, lead to three core challenges for traditional centralized prediction methods: First, multi-node heterogeneous data have time and space mismatches due to differences in grid-connected locations, and conventional prediction models are difficult to capture dynamic correlations; second, the non-stationary power fluctuations caused by sudden changes in irradiation and cloud cover make single-mode predictions have significant lag errors; third, the difference frequency timestamp data stream generated by asynchronous acquisition equipment reduces the credibility of the prediction interval. Existing technologies generally use fixed threshold filtering or single LSTM time series modeling, which can neither cope with multi-dimensional data coupling in distributed scenarios nor lack an effective decoupling mechanism for non-stationary components, resulting in the dual defects of mean shift and missing extreme values ​​in the prediction results. In this context, it is urgent to build a prediction system that integrates topology perception and modal adaptation to achieve accurate prediction of power fluctuations through data-model collaborative optimization. Summary of the invention

[0003] This application provides a distributed photovoltaic power prediction method and system based on machine learning, aiming to solve the technical problem of insufficient prediction accuracy caused by complex topological structure and asynchronous data collection of distributed photovoltaic systems.

[0004] The first aspect disclosed in the present application provides a distributed photovoltaic power prediction method based on machine learning, the method comprising: interacting with the grid-connected location nodes and installed parameters of photovoltaic base stations within the photovoltaic jurisdiction area to determine the distributed topology; mining photovoltaic power elements, using the decomposition-reconstruction-prediction of element modes as the underlying logic, developing a power prediction module in the photovoltaic system and performing convergence learning and embedding connection; connecting to a front-end acquisition card and coordinating with the photovoltaic system, and collecting a sampling matrix based on photovoltaic power elements by introducing timestamp constraints, wherein the timestamp constraints include same-frequency timestamps and difference-frequency timestamps; according to a data import interface, importing the sampling matrix into the power prediction module, combining dynamic and static methods to perform interval prediction decisions, and determining power triples, wherein the power triples include power minimum-power mean-power maximum; wherein dynamic supervision of streaming data is performed based on the distributed topology, and the streaming data includes a sampling stage and a power distribution stage.

[0005] In a possible implementation, the following processing is performed: interacting with the data source of the photovoltaic system to determine a first element source based on the photovoltaic power element; determining a second element source and configuring the front-end acquisition card, wherein the second element source is the non-first element source part of the photovoltaic power element; establishing a communication connection between the front-end acquisition card and the photovoltaic system, and establishing a data import interface with the power prediction module for the first element source and the second element source.

[0006] In a possible implementation, the following processing is performed: photovoltaic power records are retrieved to mine the photovoltaic power elements, wherein the photovoltaic power elements satisfy a preset correlation; for the photovoltaic power elements, element modal decomposition is performed to divide the elements into a stationary set and a non-stationary set, wherein the element modality includes a scene modality and a dimensional modality; the non-stationary set is decomposed and reconstructed to determine the trend component and the oscillation component as a reconstructed set; photovoltaic prediction learning is performed for the stationary set and the reconstructed set.

[0007] In a possible implementation, the following processing is performed: a first scene modality is acquired, and a first-order division is performed on the photovoltaic power element for the first scene modality to determine the first-order division element; based on the dimensional modality, a second-order decomposition is performed on the non-stationary element within the first-order division element to determine the second-order decomposition element; an association is established between the first-order division element and the second-order decomposition element, and a stationary label and a non-stationary label are identified as the modal decomposition division result of the first scene modality.

[0008] In a possible implementation, the following processing is performed: the non-stationary set is determined by performing label recognition; the non-stationary set is divided based on time series characteristics at a preset period in combination with the photovoltaic power record to determine the trend component and the oscillation component as the reconstruction set.

[0009] In a possible implementation, the following processing is performed: the power prediction module is developed and embedded assembly is performed using static prediction learning based on the stable set and dynamic prediction learning based on the reconstructed set.

[0010] In a possible implementation, the following processing is performed: determining sampling conditions, wherein the sampling conditions include a first sampling node, a second sampling node, and a third sampling node based on a same-interval time series; based on the second sampling node, performing same-frequency sampling of the stationary set based on the front-end acquisition card and the photovoltaic system to determine a first sampling matrix; based on the first sampling node-the second sampling node-the third sampling node, performing same-frequency sampling of the non-stationary set based on the front-end acquisition card and the photovoltaic system to determine a second sampling matrix; and fusing the first sampling matrix and the second sampling matrix to determine the sampling matrix.

[0011] In a possible implementation, the following processing is performed: the sampling matrix is ​​preprocessed based on the prediction of photovoltaic power elements and the decomposition-reconstruction based on the element mode to determine the preprocessing matrix; by identifying the stationary label and the non-stationary label, the static and dynamic combined prediction of the preprocessing matrix is ​​performed to determine the power triplet, wherein the decision is made by fusing the static moment based on the stationary label part and the dynamic time zone of the non-stationary label.

[0012] In a possible implementation, the following processing is performed: identifying extreme values ​​of matrix data based on non-stationary labels, wherein the extreme values ​​of the matrix data include maximum values ​​and minimum values; identifying matrix data based on stationary labels, and determining power maxima and power minima in combination with the extreme values ​​of the matrix data; averaging the power maxima and the power minima to determine a power triplet.

[0013] Another aspect disclosed in the present application provides a distributed photovoltaic power prediction system based on machine learning, the system comprising: a topology determination unit: determining the distributed topology by interacting with the grid-connected location nodes and installed parameters of photovoltaic base stations within the photovoltaic jurisdiction area; a power prediction module development unit: mining photovoltaic power elements, taking the decomposition-reconstruction-prediction of element modes as the underlying logic, developing a power prediction module in the photovoltaic system and performing convergence learning and embedding connection; a sampling matrix acquisition unit: accessing the front-end acquisition card and cooperating with the photovoltaic system, and acquiring a sampling matrix based on photovoltaic power elements by introducing timestamp constraints, wherein the timestamp constraints include same-frequency timestamps and difference-frequency timestamps; a prediction decision unit: importing the sampling matrix into the power prediction module according to the data import interface, performing interval prediction decisions by combining dynamic and static, and determining a power triplet, wherein the power triplet includes power minimum-power mean-power maximum; a dynamic supervision unit: wherein dynamic supervision of stream data is performed based on the distributed topology, and the stream data includes a sampling stage and a power distribution stage.

[0014] One or more technical solutions provided in this application have at least the following technical effects or advantages: The above-mentioned distributed photovoltaic power prediction method based on machine learning first interacts with the grid-connected location nodes and installed parameters of photovoltaic base stations in the photovoltaic jurisdiction area to determine the distributed topology, then mines photovoltaic power elements, takes the decomposition-reconstruction-prediction of element modes as the underlying logic, develops a power prediction module in the photovoltaic system and performs convergence learning and embedding, then connects to the front-end acquisition card and cooperates with the photovoltaic system, and collects the sampling matrix based on photovoltaic power elements by introducing timestamp constraints, wherein the timestamp constraints include the same frequency timestamp and the difference frequency timestamp, and then imports the sampling matrix into the power prediction module according to the data import interface, and performs interval prediction decisions by combining dynamic and static, and determines the power triplet, wherein the power triplet includes the power minimum value-power mean value-power maximum value, and finally performs dynamic supervision of stream data based on the distributed topology, and the stream data includes the sampling stage and the power distribution stage. The technical effect of high-precision and robust power prediction is achieved.

[0015] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 The figure is a flow chart of a distributed photovoltaic power prediction method based on machine learning in one embodiment.

[0018] Figure 2 The present invention is an architecture diagram of a distributed photovoltaic power prediction system based on machine learning in one embodiment.

[0019] Explanation of the reference numerals: topology determination unit 11 , power prediction module development unit 12 , sampling matrix acquisition unit 13 , prediction decision unit 14 , dynamic supervision unit 15 . DETAILED DESCRIPTION

[0020] The embodiments of the present application provide a distributed photovoltaic power prediction method and system based on machine learning to solve the technical problem of insufficient prediction accuracy of distributed photovoltaic systems due to complex topological structures and asynchronous data collection.

[0021] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0022] It should be noted that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules that are not explicitly listed or inherent to these processes, methods, products or devices.

[0023] Embodiment 1, as Figure 1 As shown, the present application provides a distributed photovoltaic power prediction method based on machine learning, the method comprising: The grid-connected location nodes and installed capacity parameters of photovoltaic base stations within the interactive photovoltaic jurisdiction area determine the distributed topology.

[0024] In the embodiment of the present application, when predicting photovoltaic power, it is necessary to clarify the distribution of photovoltaic base stations and their technical parameters in order to build an accurate distributed topology. Specifically, first, the grid-connected location nodes and installed parameters of all photovoltaic base stations in the photovoltaic jurisdiction are collected. This information is the basis for building the topology, where the grid-connected location node refers to the specific location where the photovoltaic power station is connected to the power grid; the installed parameters include installed capacity (the total power generation capacity of the photovoltaic system), operating parameters of photovoltaic modules (such as module efficiency, temperature coefficient, etc.), and photoelectric conversion characteristics (such as maximum power point voltage, current, etc.). Subsequently, according to the grid-connected location nodes of the photovoltaic base stations, their distribution in the power grid is determined. These nodes may be distributed in different geographical locations and connected to different power grid levels (such as low-voltage, medium-voltage or high-voltage power grids), thus forming a complex network structure. Afterwards, the installed parameters of each PV base station are associated with the grid-connected location nodes, and the graph neural network is used to construct a distributed topological structure of the PV jurisdiction area. This topological structure reflects the physical distribution of the PV base stations, the grid connection relationship and the technical characteristics of each node. It will be dynamically updated every 15 minutes (which can be self-regulated) to accurately reflect the real-time distribution of the PV cluster in the distribution network, thereby providing the necessary space and parameter basis for subsequent power forecasting.

[0025] By mining photovoltaic power elements and taking the decomposition-reconstruction-prediction of element modes as the underlying logic, a power prediction module is developed in the photovoltaic system and convergent learning and embedded connection are performed.

[0026] In one embodiment, in photovoltaic power prediction, in order to achieve accurate prediction of power fluctuations, it is necessary to deeply explore the elements of photovoltaic power and develop a power prediction module based on the underlying logic of decomposition-reconstruction-prediction. Specifically, photovoltaic power is affected by many factors, including irradiation intensity, temperature, cloud cover, component efficiency, etc. By analyzing the photovoltaic power records, the influencing factors with high correlation can be extracted as photovoltaic power elements, such as irradiation intensity, temperature change, etc. Subsequently, the photovoltaic power elements determined by mining are decomposed in multiple stages to determine multiple modal components, such as scene modalities and dimensional modalities. By combining scene modalities and dimensional modalities, the photovoltaic power elements are divided into stationary sets and non-stationary sets to capture different features respectively. Afterwards, the decomposed non-stationary set is divided based on the time series characteristics, noise or redundant information is removed, and components that contribute significantly to the prediction are retained, thereby completing the reconstruction of the non-stationary set and obtaining the reconstructed set. Then, based on the stable set and the reconstruction set, the prediction model training and learning are carried out respectively until the model converges, and then the power prediction module is developed by connecting the models in series to more accurately capture the dynamic characteristics of power fluctuations. Finally, this power prediction module is embedded in the control or scheduling platform of the photovoltaic system, and the non-stationarity and complexity of photovoltaic power are dealt with by real-time processing of photovoltaic data, thereby providing reliable support for the stable operation of the photovoltaic system.

[0027] Furthermore, the present application provides a power prediction module developed in a photovoltaic system based on the decomposition-reconstruction-prediction of element modes as the underlying logic, including: Retrieve photovoltaic power records and mine the photovoltaic power elements, wherein the photovoltaic power elements meet the preset correlation; perform element modal decomposition on the photovoltaic power elements and divide them into a stationary set and a non-stationary set, wherein the element modality includes scene modality and dimensional modality; perform element decomposition and reconstruct the non-stationary set to determine the trend component and the oscillation component as the reconstructed set; perform photovoltaic prediction learning on the stationary set and the reconstructed set.

[0028] Preferably, historical photovoltaic power records are obtained from the photovoltaic system, including power data in time series and related environmental data (such as light, temperature, humidity, wind speed, etc.), and the historical photovoltaic power records are normalized by the maximum-minimum method so that these data are in the same dimension. Then, based on the normalized historical photovoltaic power records, the Pearson correlation coefficient is used to calculate the correlation between each environmental data and the photovoltaic power, and the calculation result is compared with the preset correlation to screen out environmental data that are highly correlated with photovoltaic power fluctuations to form photovoltaic power elements; then, the photovoltaic power elements are divided into a first-order and a second-order division according to the current weather conditions, that is, according to the current weather conditions and element dimensions, the photovoltaic power elements are divided into a stationary set and a non-stationary set, wherein the stationary set refers to a set of elements that show stability and regularity in the time series, the change trend of these elements is relatively gentle, the fluctuation is small, and they usually have periodic or trend characteristics, such as the irradiation intensity on a sunny day; the non-stationary set refers to a set of elements that show significant fluctuations or irregular changes in the time series, the change trend of these elements is complex, the fluctuation is large, and they may be strongly affected by external factors, such as the irradiation intensity on a cloudy day. Afterwards, the trend component and the oscillation component are determined by dividing the non-stationary set based on the time series characteristics, where the trend component reflects the long-term change trend of power; the oscillation component reflects the short-term fluctuation of power. The trend component and the oscillation component are combined as the reconstruction result of the non-stationary set, that is, the reconstruction set. Then, the stationary set is statically predicted and learned using traditional time series models (such as ARIMA, SARIMA) or machine learning models (such as LSTM, support vector machine), and the reconstruction set is dynamically predicted and learned using machine learning models (such as LSTM, random forest), thereby constructing two prediction models, one for static prediction based on the stationary set and the other for dynamic prediction based on the reconstruction set. These two models are connected in series to form a power prediction module, which is embedded in the control or dispatching platform of the photovoltaic system to achieve accurate prediction of photovoltaic power, effectively deal with the complexity and non-stationarity of power fluctuations, and provide reliable support for the stable operation of the photovoltaic system.

[0029] Furthermore, the present application provides a method for performing factor modal decomposition and dividing a stationary set and a non-stationary set, including: The first scene modality is obtained, and the photovoltaic power element is divided into the first order according to the first scene modality to determine the first order division element; based on the dimensional modality, the non-stationary element within the first order division element is decomposed into the second order to determine the second order decomposition element; the association between the first order division element and the second order decomposition element is established, and the stationary label and the non-stationary label are identified as the modal decomposition division result of the first scene modality.

[0030] Optionally, the current weather conditions (such as sunny, cloudy, rainy, etc.) are obtained by online query, and the weather is used as the first scene mode as the basis for subsequent analysis. Under the first scene mode, the photovoltaic power element is divided into the first order. In this process, the first scene mode is used as an index to extract the historical data with the same weather as the current weather from the historical photovoltaic power record to form the first historical photovoltaic power record. Then, the data of each photovoltaic power element in the first historical photovoltaic power record is divided according to the preset time step to obtain multiple sub-intervals of each photovoltaic power element, and the mean and variance of each sub-interval are calculated. Then, the ADF statistic and p-value are calculated based on the data of each photovoltaic power element. If the p-value is less than the significance level, and the mean and variance change between different sub-intervals is less than or equal to the stability threshold, it means that the photovoltaic power element is relatively stable. At this time, the photovoltaic power element will be added to the stable set. On the contrary, if the p-value is greater than or equal to the significance level, and the mean and variance change between different sub-intervals is greater than the stability threshold, it means that the photovoltaic power element fluctuates greatly. At this time, the photovoltaic power element will be added to the non-stationary set. The two data sets will jointly form the first-order division element. Subsequently, the non-stationary elements in the first-order partition are decomposed in the second order based on the dimensional modality, wherein the dimensional modality is the sub-dimension corresponding to the different elements set in advance according to the relevant field knowledge, for example, the sub-dimension corresponding to the illumination is the illumination intensity, illumination angle, spectral distribution, etc., and the sub-dimension of the wind speed is the wind speed magnitude, direction, wind speed volatility, etc. Through this dimensional modality, the sub-dimension data of the non-stationary elements in the non-stationary set can be extracted, that is, the data of each sub-dimension is extracted from the first historical photovoltaic power record, and then the same stability analysis is performed on the data of each sub-dimension, and the sub-dimension that causes the non-stationary elements in the first-order partition elements to fluctuate is determined, and the determined sub-dimension is used as the second-order decomposition element. Afterwards, each first-order partition element (stationary and non-stationary) is associated with the corresponding second-order decomposition element, for example, the stationary element is directly associated with itself, and the non-stationary element is associated with its decomposed sub-dimension, and then each element is identified, and the identification content includes a stationary label and a non-stationary label. After the identification is completed, a hierarchical modal decomposition structure is formed as the modal decomposition result of the first scene modal. This modal decomposition result can be represented as a tree structure, in which the root node is the first scene modal, the child nodes are the first-order decomposition elements, and the leaf nodes are the second-order decomposition elements. This modal decomposition result provides a clearer data structure for subsequent modeling and prediction, and can improve the accuracy and robustness of the model.

[0031] Furthermore, the present application provides a method for reconstructing the non-stationary set by decomposing the non-stationary set into elements, including: The non-stationary set is determined by performing label recognition; the non-stationary set is divided based on time series characteristics in combination with the photovoltaic power record at a preset period to determine a trend component and an oscillation component as the reconstruction set.

[0032] Optionally, according to the labels in the modal decomposition partitioning results, data with non-stationary labels are extracted from them to determine the non-stationary set. Subsequently, according to the characteristics of photovoltaic power (such as daily cycle, seasonal cycle, etc.), the non-stationary set is divided into multiple period segments, for example, by day, by hour or by season, and then a smoothing method (such as moving average, low-pass filtering) is used to extract the trend component, which reflects the long-term change law of the data. The original non-stationary set is subtracted from the trend component to obtain the oscillation component, which reflects the short-term fluctuation characteristics of the data. Taking the moving average as an example, the preset period is used as the moving window size, and the average value of the first n data points is calculated for each time point as the trend component. For the first n-1 data points, the complete moving average cannot be calculated, and you can choose to directly retain the original value or use other methods (such as filling, interpolation) for processing. Afterwards, for each time point, the trend component is subtracted from the original value at the current moment to obtain the oscillation component at that time point, and then all trend components and oscillation components are summarized to obtain a reconstructed set for subsequent analysis and modeling.

[0033] Furthermore, the present application provides a method of developing the power prediction module and performing embedded assembly based on static prediction learning based on the stable set and dynamic prediction learning based on the reconstructed set.

[0034] Optionally, after obtaining the stationary set and the reconstruction set, these two sets of data will be used to develop a power prediction module. Specifically, taking LSTM as an example, a framework of a static prediction model and a dynamic prediction model is constructed based on LSTM. Both models include an input layer, an LSTM layer, a fully connected layer, and an output layer. The output layer of the dynamic prediction model and the input layer of the static prediction model are then connected in series to form a power prediction module, and the weights of the dynamic prediction model and the static prediction model are initialized using random numbers and other methods. Subsequently, the reconstruction set and the stationary set are input into the power prediction module together, and the dynamic prediction model is dynamically predicted and learned using the reconstruction set, that is, the data corresponding to the photovoltaic power element in the reconstruction set is passed as input data to the initialized dynamic prediction model for forward propagation, and is passed layer by layer through the input layer, LSTM layer, fully connected layer, and output layer to calculate the prediction results including the initial trend component and the initial oscillation component within a preset period, and then each group of initial trend components and initial oscillation components are added through the output layer of the power prediction module to obtain multiple predicted power values ​​within a preset period, and the maximum and minimum values ​​of the power are extracted from them to form the matrix data extreme value. After that, the obtained matrix data extreme values ​​and stable sets are passed to the static prediction model for static prediction learning, that is, the data corresponding to the photovoltaic power elements in the matrix data extreme values ​​and stable sets are passed as input data to the initialized static prediction model for forward propagation, and the trend component and oscillation component within a preset period are calculated. Then, the mean square error loss function is used to calculate the loss value between the prediction result and the label data (the power data is divided based on the time series characteristics using the same method as mentioned above, and the trend component and oscillation component of the power are obtained), and the gradient of the loss to the weights of each layer of the two models is calculated layer by layer through back propagation. Then, the Adam optimizer is used to optimize the two model parameters, and the weights are adjusted to minimize the value of the loss function. The above process is repeated until the maximum number of iterations is reached. After the training is completed, the model performance is tested using data that is not used for training, and the accuracy of the model on the regression task of the trend component and oscillation component within a preset period is evaluated. If the accuracy meets the expected accuracy, the parameters of the current power prediction module are saved. Otherwise, the hyperparameters such as the learning rate and the number of training batches are adjusted to further improve the prediction effect of the power prediction module. Finally, the determined power prediction module is embedded into the control or scheduling platform of the photovoltaic system to achieve accurate prediction of photovoltaic power.

[0035] The front-end acquisition card is connected and cooperates with the photovoltaic system to collect a sampling matrix based on photovoltaic power elements by introducing a timestamp constraint, wherein the timestamp constraint includes a same-frequency timestamp and a difference-frequency timestamp.

[0036] In one embodiment, when performing real-time acquisition of photovoltaic power elements, it is first necessary to access the front-end acquisition card, which is a hardware device used to collect environmental information such as light intensity, temperature, humidity, etc. from sensors, instruments or other data sources in real time. By connecting the front-end acquisition card to the photovoltaic system, it can be ensured that the environmental data can be accurately and stably transmitted. Subsequently, a timestamp constraint is introduced to accurately control the time of data acquisition to ensure the synchronization and consistency of the data. This timestamp constraint includes a same-frequency timestamp and a difference-frequency timestamp, wherein the same-frequency timestamp data means that the time interval of data acquisition is fixed, that is, sampling is performed according to the same interval time series, for example, data is collected every 5 minutes to ensure that the time interval between each sampling point is the same; the difference-frequency timestamp means that the time interval of data acquisition is not fixed, but is dynamically adjusted according to the characteristics of the non-stationary data set. For example, in the case of drastic environmental changes or unstable photovoltaic system state, the acquisition frequency may increase or decrease to capture more detailed or more extensive changes. Afterwards, based on this timestamp constraint, multiple data sampling is performed based on the front-end acquisition card and the photovoltaic system, and then these data are fused to obtain a sampling matrix of photovoltaic power elements. Each row of the matrix represents data at a time point, and each column represents a photovoltaic power element (such as light intensity, temperature, humidity, etc.). In this way, the data can be analyzed and processed more conveniently, laying the foundation for the prediction of photovoltaic power.

[0037] Furthermore, the present application provides a sampling matrix for collecting photovoltaic power elements, including: Determine sampling conditions, wherein the sampling conditions include a first sampling node, a second sampling node, and a third sampling node based on a same-interval time series; according to the second sampling node, perform same-frequency sampling of the stationary set based on the front-end acquisition card and the photovoltaic system to determine a first sampling matrix; according to the first sampling node-the second sampling node-the third sampling node, perform same-frequency sampling of the non-stationary set based on the front-end acquisition card and the photovoltaic system to determine a second sampling matrix; and fuse the first sampling matrix and the second sampling matrix to determine the sampling matrix.

[0038] Optionally, before data collection, the sampling conditions are first determined, and three sampling nodes are defined in the sampling conditions, namely, the first sampling node, the second sampling node, and the third sampling node. These nodes are determined based on the same interval time series, that is, the time intervals between them are fixed. Subsequently, the second sampling node is used as a reference, and the same frequency sampling is performed according to the photovoltaic power elements in the stable set, and then the collected stable set data is sorted into a matrix form according to the time series to form a first sampling matrix, and each row of the first sampling matrix corresponds to a sampling point, and each column corresponds to a photovoltaic power element in the stable set. Similarly, the first sampling node, the second sampling node, and the third sampling node are also used as references (starting from the first sampling node, data is collected every fixed time interval until the third sampling node), and the photovoltaic power elements in the non-stationary set are sampled at the same frequency, and then the collected non-stationary set data is sorted into a matrix form according to the time series to form a second sampling matrix, and each row of the second sampling matrix corresponds to a sampling point, and each column corresponds to a photovoltaic power element in the non-stationary set. These two acquisition methods correspond to the moment prediction of the stationary set and the time zone dynamic prediction of the non-stationary set, respectively. This is because the data of the stationary set changes relatively slowly, which is suitable for sampling at fixed time intervals, and only the data at the current moment needs to be paid attention to during prediction, while the data of the non-stationary set changes dramatically or has anomalies, and it is necessary to capture the dynamic change trend in the time series. After that, the first sampling matrix and the second sampling matrix are aligned according to the timestamp to ensure the consistency of the time series of the data. After alignment, the data of the two matrices are merged to generate the final sampling matrix. In summary, through the above acquisition method, the continuity of the data is guaranteed, and the anomalies or drastic fluctuations in environmental changes can be captured, providing a reliable data basis for subsequent photovoltaic power prediction.

[0039] According to the data import interface, the sampling matrix is ​​imported into the power prediction module, and interval prediction decisions are made by combining dynamic and static methods to determine the power triplet, wherein the power triplet includes the power minimum value-power mean value-power maximum value.

[0040] In one embodiment, after the sampling matrix is ​​collected, the sampling matrix is ​​transmitted to the power prediction module through a pre-established data import interface. The sampling matrix contains data of a stationary set and a non-stationary set, and each row corresponds to an environmental factor (such as light intensity, temperature, humidity, etc.) at a time point. Subsequently, the sampling matrix is ​​preprocessed, and then the power prediction module is used to make interval prediction decisions based on the results of the preprocessing by combining a dynamic prediction model with a static prediction model to determine a power triplet. This power triplet includes power minimum-power mean-power maximum. This interval prediction decision method can provide a more comprehensive and reliable basis for the operation and management of photovoltaic systems.

[0041] Furthermore, the present application also provides that before the sampling matrix is ​​imported into the power prediction module, the steps include: Interact with the data source of the photovoltaic system to determine a first element source based on the photovoltaic power element; determine a second element source and configure the front-end acquisition card, wherein the second element source is the non-first element source portion of the photovoltaic power element; establish a communication connection between the front-end acquisition card and the photovoltaic system, and establish a data import interface with the power prediction module for the first element source and the second element source.

[0042] Optionally, in the process of interacting with the data source of the photovoltaic system and establishing a data import interface, it is necessary to first analyze the photovoltaic power factor and determine the first factor source based on the photovoltaic power factor. This first factor source refers to the environmental factor data directly collected by sensors or instruments, such as light intensity, temperature, humidity, wind speed, etc. These data are obtained in real time by sensors or data acquisition equipment deployed in the photovoltaic system and belong to local direct measurement data. Subsequently, the second factor source is determined, that is, the other part of the photovoltaic power factor except the first factor source. This second factor source refers to the environmental factor data obtained through online query or external data interface. These data cannot be directly obtained by the first factor source, such as atmospheric pressure, cloud coverage, air quality index (AQI), etc. These data usually come from meteorological stations, satellite remote sensing or other external data services, and need to be obtained through online means. Afterwards, configure a front-end acquisition card to collect data from the first and second element sources. This front-end acquisition card obtains data from the second element source from an external data service through a network interface or API, and is physically connected to the sensors and data acquisition equipment of the photovoltaic system to ensure that the data from the first element source can be correctly transmitted. Configure the communication parameters of the front-end acquisition card according to the communication protocol of the photovoltaic system (such as Modbus, CAN, etc.), establish a data import interface, and ensure that the data from the first and second element sources can be correctly transmitted to the power prediction module embedded in the photovoltaic system, thereby providing reliable data support for subsequent photovoltaic power prediction.

[0043] Furthermore, the present application also provides the determination of the power triplet, including: The sampling matrix is ​​preprocessed based on the prediction of photovoltaic power elements and the decomposition-reconstruction based on the element mode to determine the preprocessing matrix; by identifying the stationary label and the non-stationary label, the preprocessing matrix is ​​subjected to the dynamic and static combined prediction to determine the power triplet, wherein the decision is made by fusing the static moment based on the stationary label part and the dynamic time zone of the non-stationary label.

[0044] Optionally, after obtaining the sampling matrix, the photovoltaic power element is used as the basis for prediction, and the data corresponding to the photovoltaic power element is extracted from the sampling matrix, and the extracted data is preprocessed based on the same steps of decomposition-reconstruction of the aforementioned element mode to obtain a preprocessing matrix, which includes all the data in the current stationary set and the reconstruction set. Subsequently, by identifying stationary labels and non-stationary labels, the stationary set and the reconstruction set are extracted from the preprocessing matrix, and the stationary set and the reconstruction set are transmitted to the power prediction module for dynamic and static combined prediction. In this process, the dynamic prediction model in the power prediction module is used to make dynamic time zone decisions on the reconstruction set to obtain the extreme value of the matrix data, and then the stationary set and the extreme value of the matrix data are combined, and the static prediction model is used to make static time decisions to obtain the power triple. In summary, this method of determining the power triple through dynamic and static combined prediction can make full use of the static and dynamic characteristics of the data, thereby improving the accuracy and robustness of photovoltaic power prediction.

[0045] Furthermore, the present application also provides a method for making a decision by fusing the static moment based on the stationary label part with the dynamic time zone of the non-stationary label, including: Identify extreme values ​​of matrix data based on non-stationary labels, wherein the extreme values ​​of the matrix data include maximum values ​​and minimum values; identify matrix data based on stationary labels, and determine power maximum values ​​and power minimum values ​​in combination with the extreme values ​​of the matrix data; average the power maximum values ​​and the power minimum values ​​to determine a power triplet.

[0046] Optionally, after obtaining the extreme value of the matrix data of the non-stationary label through the dynamic prediction model, this matrix data extreme value will be used as input data together with the data in the stationary set, wherein the matrix data extreme value includes the maximum and minimum values ​​of the initial power, that is, only the maximum and minimum values ​​of the photovoltaic power elements with non-stationary labels are considered. Subsequently, this input data is input into the static prediction model for static moment decision-making to obtain the power maximum and power minimum values, which are the results obtained by combining all photovoltaic power elements and have high accuracy. Afterwards, the power maximum and power minimum values ​​are averaged, and the calculation results are combined with the power maximum and power minimum values ​​to form a power triplet. This prediction method makes full use of the static and dynamic characteristics of the data, can more accurately capture the fluctuation range of photovoltaic power, and provide a more comprehensive and reliable basis for the operation and management of photovoltaic systems.

[0047] Wherein, dynamic supervision of flow data is performed based on the distributed topology, and the flow data includes a sampling phase and a power distribution phase.

[0048] In one embodiment, the streaming data is dynamically supervised under the distributed topology, mainly involving the sampling stage and the power distribution stage. In the sampling stage, each node in the distributed photovoltaic system will collect photovoltaic power element data and power data in real time, and transmit the photovoltaic power element data to the photovoltaic system through the distributed network for prediction. In the power distribution stage, based on the data obtained in the sampling stage, the distribution of photovoltaic power will be dynamically analyzed and supervised, that is, the future power distribution will be predicted based on the photovoltaic power data, and then the actual power output of each node will be monitored in real time, and compared with the prediction results, to identify abnormal situations (whether the degree of deviation exceeds the tolerance range), such as power sag or surge, and take corresponding control measures (such as adjusting the inverter output or redistributing the load). Through the collaborative work of these two stages, the distributed topology can achieve comprehensive dynamic supervision of the streaming data of the photovoltaic system, ensuring the efficient operation and reliable power supply of the photovoltaic system.

[0049] In summary, the embodiments of the present application have at least the following technical effects: The embodiment of the present application interacts with the grid-connected location nodes and installed parameters of photovoltaic base stations in the photovoltaic jurisdiction area to determine the distributed topology; excavates photovoltaic power elements, takes the decomposition-reconstruction-prediction of element modes as the underlying logic, develops a power prediction module in the photovoltaic system and performs convergence learning and embedded connection; accesses the front-end acquisition card and cooperates with the photovoltaic system, and collects the sampling matrix based on the photovoltaic power element by introducing timestamp constraints, wherein the timestamp constraints include the same frequency timestamp and the difference frequency timestamp; according to the data import interface, the sampling matrix is ​​imported into the power prediction module, and the interval prediction decision is made by combining dynamic and static, and the power triple is determined, wherein the power triple includes the power minimum value-power mean value-power maximum value; wherein, the streaming data is dynamically supervised based on the distributed topology, and the streaming data includes the sampling stage and the power distribution stage. These technical effects jointly solve the technical problem of insufficient prediction accuracy of distributed photovoltaic systems due to complex topological structures and asynchronous data acquisition, and realize the technical effect of high-precision and robust power prediction through distributed topology construction, modal decomposition-reconstruction prediction mechanism and dynamic and static combined interval prediction decision.

[0050] Embodiment 2 is based on the same inventive concept as the distributed photovoltaic power prediction method based on machine learning in the above embodiment. Figure 2 As shown, the present application provides a distributed photovoltaic power prediction system based on machine learning, the system comprising: A topology determination unit 11: The grid-connected location nodes and installed parameters of the photovoltaic base stations in the interactive photovoltaic jurisdiction area are used to determine the distributed topology; a power prediction module development unit 12: Exploring photovoltaic power elements, using the decomposition-reconstruction-prediction of element modes as the underlying logic, developing a power prediction module in the photovoltaic system and performing convergence learning and embedding connection; a sampling matrix acquisition unit 13: Connecting to the front-end acquisition card and coordinating with the photovoltaic system, by introducing timestamp constraints, acquiring a sampling matrix based on photovoltaic power elements, wherein the timestamp constraints include the same frequency timestamp and the difference frequency timestamp; a prediction decision unit 14: According to the data import interface, the sampling matrix is ​​imported into the power prediction module, and interval prediction decisions are performed by combining dynamic and static methods to determine the power triplet, wherein the power triplet includes the power minimum value-power mean value-power maximum value; a dynamic supervision unit 15: wherein, dynamic supervision of stream data is performed based on the distributed topology, and the stream data includes a sampling stage and a power distribution stage.

[0051] Furthermore, the power prediction module development unit 12 is also used to perform the following method: Retrieve photovoltaic power records and mine the photovoltaic power elements, wherein the photovoltaic power elements meet the preset correlation; perform element modal decomposition on the photovoltaic power elements and divide them into a stationary set and a non-stationary set, wherein the element modality includes scene modality and dimensional modality; perform element decomposition and reconstruct the non-stationary set to determine the trend component and the oscillation component as the reconstructed set; perform photovoltaic prediction learning on the stationary set and the reconstructed set.

[0052] Furthermore, the power prediction module development unit 12 is also used to perform the following method: The first scene modality is obtained, and the photovoltaic power element is divided into the first order according to the first scene modality to determine the first order division element; based on the dimensional modality, the non-stationary element within the first order division element is decomposed into the second order to determine the second order decomposition element; the association between the first order division element and the second order decomposition element is established, and the stationary label and the non-stationary label are identified as the modal decomposition division result of the first scene modality.

[0053] Furthermore, the power prediction module development unit 12 is also used to perform the following method: The non-stationary set is determined by performing label recognition; the non-stationary set is divided based on time series characteristics in combination with the photovoltaic power record at a preset period to determine a trend component and an oscillation component as the reconstruction set.

[0054] Furthermore, the power prediction module development unit 12 is also used to perform the following method: The power prediction module is developed and embedded assembly is performed by static prediction learning based on the stable set and dynamic prediction learning based on the reconstructed set.

[0055] Furthermore, the sampling matrix acquisition unit 13 is also used to perform the following method: Determine sampling conditions, wherein the sampling conditions include a first sampling node, a second sampling node, and a third sampling node based on a same-interval time series; according to the second sampling node, perform same-frequency sampling of the stationary set based on the front-end acquisition card and the photovoltaic system to determine a first sampling matrix; according to the first sampling node-the second sampling node-the third sampling node, perform same-frequency sampling of the non-stationary set based on the front-end acquisition card and the photovoltaic system to determine a second sampling matrix; and fuse the first sampling matrix and the second sampling matrix to determine the sampling matrix.

[0056] Furthermore, the prediction decision unit 14 is also used to execute the following method: Interact with the data source of the photovoltaic system to determine a first element source based on the photovoltaic power element; determine a second element source and configure the front-end acquisition card, wherein the second element source is the non-first element source portion of the photovoltaic power element; establish a communication connection between the front-end acquisition card and the photovoltaic system, and establish a data import interface with the power prediction module for the first element source and the second element source.

[0057] Furthermore, the prediction decision unit 14 is also used to execute the following method: The sampling matrix is ​​preprocessed based on the prediction of photovoltaic power elements and the decomposition-reconstruction based on the element mode to determine the preprocessing matrix; by identifying the stationary label and the non-stationary label, the preprocessing matrix is ​​subjected to the dynamic and static combined prediction to determine the power triplet, wherein the decision is made by fusing the static moment based on the stationary label part and the dynamic time zone of the non-stationary label.

[0058] Furthermore, the prediction decision unit 14 is also used to execute the following method: Identify extreme values ​​of matrix data based on non-stationary labels, wherein the extreme values ​​of the matrix data include maximum values ​​and minimum values; identify matrix data based on stationary labels, and determine power maximum values ​​and power minimum values ​​in combination with the extreme values ​​of the matrix data; average the power maximum values ​​and the power minimum values ​​to determine a power triplet.

[0059] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of this specification are described. The processes depicted in the accompanying drawings do not necessarily require the specific order and continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0060] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0061] This specification and drawings are merely exemplary illustrations of the present application and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, a person skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application intends to include these modifications and variations.

Claims

1. A distributed photovoltaic power prediction method based on machine learning, characterized in that: The method comprises: Interactive PV determines the grid-connected location nodes and installed capacity parameters of PV base stations within the jurisdiction of the PV system, and determines the distributed topology; Mining photovoltaic power elements, using the decomposition-reconstruction-prediction of element modalities as the underlying logic, developing a power prediction module in the photovoltaic system and performing convergent learning and embedded connection; Accessing the front-end acquisition card and cooperating with the photovoltaic system, collecting a sampling matrix based on photovoltaic power elements by introducing a timestamp constraint, wherein the timestamp constraint includes a same-frequency timestamp and a difference-frequency timestamp; According to the data import interface, the sampling matrix is ​​imported into the power prediction module, and interval prediction decisions are made by combining dynamic and static methods to determine a power triplet, wherein the power triplet includes a power minimum value, a power mean value, and a power maximum value; Wherein, dynamic supervision of flow data is performed based on the distributed topology, and the flow data includes a sampling phase and a power distribution phase.

2. A distributed photovoltaic power prediction method based on machine learning as claimed in claim 1, characterized in that: Before the sampling matrix is ​​imported into the power prediction module, the method includes: Interacting with the data source of the photovoltaic system to determine a first element source based on the photovoltaic power element; Determine a second element source and configure the front-end acquisition card, wherein the second element source is a non-first element source portion of the photovoltaic power element; A communication connection is established between the front-end acquisition card and the photovoltaic system, and a data import interface is established with the power prediction module for the first element source and the second element source.

3. A distributed photovoltaic power prediction method based on machine learning as claimed in claim 1, characterized in that: Based on the decomposition-reconstruction-prediction of element modes as the underlying logic, a power prediction module is developed in the photovoltaic system, including: Retrieving photovoltaic power records and mining the photovoltaic power elements, wherein the photovoltaic power elements satisfy a preset correlation; For the photovoltaic power factor, factor modal decomposition is performed to divide the photovoltaic power factor into a stationary set and a non-stationary set, wherein the factor modality includes a scene modality and a dimensional modality. Decomposing and reconstructing the non-stationary set, and determining a trend component and an oscillation component as a reconstructed set; Photovoltaic prediction learning is performed on the stable set and the reconstructed set.

4. A distributed photovoltaic power prediction method based on machine learning as claimed in claim 3, characterized in that: Perform factor modal decomposition and divide the stationary set and non-stationary set, including: Acquire a first scene modality, perform a first-order division on the photovoltaic power element according to the first scene modality, and determine a first-order division element; Based on the dimensional mode, performing a second-order decomposition on the non-stationary elements within the first-order division elements to determine the second-order decomposition elements; An association is established between the first-order division element and the second-order decomposition element, and a stationary label and a non-stationary label are identified as a modal decomposition division result of the first scene modality.

5. A distributed photovoltaic power prediction method based on machine learning as claimed in claim 4, characterized in that: Decomposing and reconstructing the non-stationary set into elements includes: Determine the non-stationary set by performing label identification; In combination with the photovoltaic power record at a preset period, the non-stationary set is divided based on time series characteristics to determine a trend component and an oscillation component as the reconstruction set.

6. A distributed photovoltaic power prediction method based on machine learning as claimed in claim 5, characterized in that: The power prediction module is developed and embedded assembly is performed by static prediction learning based on the stable set and dynamic prediction learning based on the reconstructed set.

7. A distributed photovoltaic power prediction method based on machine learning as claimed in claim 3, characterized in that: Acquisition is based on a sampling matrix of photovoltaic power elements, including: Determine sampling conditions, wherein the sampling conditions include a first sampling node, a second sampling node, and a third sampling node based on a time series with the same interval; According to the second sampling node, the stationary set is sampled based on the same frequency of the front-end acquisition card and the photovoltaic system to determine a first sampling matrix; According to the first sampling node-the second sampling node-the third sampling node, the non-stationary set is sampled based on the same frequency of the front-end acquisition card and the photovoltaic system to determine a second sampling matrix; The first sampling matrix and the second sampling matrix are fused to determine the sampling matrix.

8. The distributed photovoltaic power prediction method based on machine learning according to claim 1, characterized in that: The determining of the power triplet comprises: Preprocessing the sampling matrix based on photovoltaic power factor prediction and factor mode-based decomposition-reconstruction to determine a preprocessing matrix; By identifying stationary labels and non-stationary labels, a combined dynamic and static prediction is performed on the preprocessing matrix to determine a power triplet, wherein a decision is made by fusing the static moment based on the stationary label part and the dynamic time zone of the non-stationary label.

9. A distributed photovoltaic power prediction method based on machine learning as claimed in claim 8, characterized in that: Decisions are made by fusing the static moments based on the stationary label part with the dynamic time zones of the non-stationary label, including: Identifying extreme values ​​of matrix data based on non-stationary labels, wherein the extreme values ​​of the matrix data include maximum values ​​and minimum values; Identify matrix data based on stationary labels, and determine power maxima and power minima in combination with extreme values ​​of the matrix data; The power maximum value and the power minimum value are averaged to determine a power triplet.

10. A distributed photovoltaic power prediction system based on machine learning, characterized in that: The system is used to execute the distributed photovoltaic power prediction method based on machine learning as described in any one of claims 1 to 9, comprising: Topology determination unit: Interact with the grid-connected location nodes and installed capacity parameters of photovoltaic base stations within the photovoltaic jurisdiction area to determine the distributed topology; Power prediction module development unit: mining photovoltaic power elements, using the decomposition-reconstruction-prediction of element modalities as the underlying logic, developing power prediction modules in photovoltaic systems and performing convergent learning and embedded connections; Sampling matrix acquisition unit: connected to the front-end acquisition card and cooperated with the photovoltaic system, and acquired the sampling matrix based on the photovoltaic power element by introducing timestamp constraints, wherein the timestamp constraints include the same frequency timestamp and the difference frequency timestamp; Prediction and decision unit: according to the data import interface, the sampling matrix is ​​imported into the power prediction module, and the dynamic and static interval prediction and decision are combined to determine the power triplet, wherein the power triplet includes the power minimum value-power mean value-power maximum value; Dynamic supervision unit: wherein, based on the distributed topology, dynamic supervision of flow data is performed, and the flow data includes a sampling phase and a power distribution phase.

Images