Distributed photovoltaic power ultra-short-term prediction method based on multi-source data fusion

Abstract

The invention belongs to the technical field of distributed photovoltaic power prediction and power distribution network operation, and discloses a distributed photovoltaic power ultra-short-term prediction method based on multi-source data fusion. The method comprises the following steps: collecting and aligning historical power and electric parameters of a photovoltaic station, meteorological monitoring, short-term and imminent forecasting, voltage of a power distribution network, power flow and states of a voltage regulating device under a unified time dimension; based on a photovoltaic output mechanism model and a linearized power flow sensitivity model or an equivalent electrical distance model, constructing a power constraint interval and a baseline power track which meet voltage and line thermal stability constraints; and through residual normalization, weighted adjacent matrix and space-time recursion prediction model modeling and prediction of a multi-station residual field, parameter increment is updated in combination with a data quality index and an error threshold, and distributed photovoltaic cluster ultra-short-term power prediction giving consideration to power distribution network security constraints and data quality is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of distributed photovoltaic power prediction and power distribution network operation, and particularly relates to a distributed photovoltaic power ultra-short-term prediction method based on multi-source data fusion. BACKGROUND

[0002] With the increasing access capacity of distributed photovoltaics in the power distribution network, photovoltaic output is affected by factors such as cloud cover, irradiance changes and load fluctuations on the 5-30 minute scale, and presents strong randomness and volatility. The dependence of power distribution network voltage control and active power balance on ultra-short-term power prediction results is increasing.

[0003] In existing projects, most methods are based on a small amount of measured data such as single-station historical active power, irradiance and temperature, and use time series models or machine learning models to predict each photovoltaic station separately. Generally, photovoltaic output is regarded as an independent power source weakly coupled with the topology of the power distribution network, node voltage and branch power flow, and the constraint conditions such as node voltage overrun and line thermal stability are rarely explicitly considered during the prediction stage. Generally, whether the operation requirements are met is checked through power flow after the prediction is completed. Actual operation shows that under the condition of high penetration rate of feeder lines, only relying on such prediction results to arrange output is prone to the situation that the predicted value itself is reasonable in numerical value, but the corresponding operating point has locally high node voltage or some branches close to the thermal stability limit.

[0004] There are also methods in engineering to correct the power curve according to short-term weather forecasts, or to set a fixed upper limit or an empirical reduction coefficient to truncate the theoretical output. However, the relationship between such upper limit and real-time node voltage, branch power flow, voltage regulating device state and electrical distance between multiple sites is not clear enough, and it is difficult to reflect the power distribution network working conditions and multi-site coupling characteristics in a timely manner. At the same time, the data collected by photovoltaic field monitoring systems, weather monitoring systems and power distribution network automation systems come from various sources, and there are generally problems such as missing, abnormal and time asynchronization. Existing prediction models generally consider input data to be complete, reliable and already aligned, and lack a mechanism for evaluating and weighting different data sources and different sites based on data quality indicators. Model parameters are usually fixed in long-term operation, and prediction results are easily affected by abnormal data and working condition changes. SUMMARY

[0005] The purpose of the present application is to provide a distributed photovoltaic power ultra-short-term prediction method based on multi-source data fusion to solve the problems raised in the background.

[0006] In order to achieve the above object, the present application provides the following technical scheme: a distributed photovoltaic power ultra-short-term prediction method based on multi-source data fusion, which is suitable for a plurality of distributed photovoltaic sites connected to a distribution network. The method combines the output mechanism of photovoltaic components with power flow constraints to construct a power baseline under the premise of distribution network safety constraints, and then models and predicts the baseline deviation in space and time to generate an ultra-short-term power prediction result under the unified consideration of photovoltaic side, meteorological side and grid side operation information and data quality differences, thereby providing input data matching the actual constraint conditions for distribution network operation and dispatching; During operation, the method first collects and aligns multi-source operation data in a unified time dimension for each prediction period, the multi-source data including: historical active power measurement values of each distributed photovoltaic site, component side current and voltage, inverter operation state and limit state, meteorological monitoring data of irradiance, ambient temperature and wind speed near the photovoltaic site, short-term weather forecast data of the corresponding area, and node voltage, branch power flow, switch and voltage regulator operation state of the distribution network side, and each node load prediction data. By aligning the multi-source data in a unified time dimension, the measurement information of the photovoltaic side, the meteorological side and the grid side are corresponded on the same time axis, thereby reducing the influence of time deviation between different data sources on subsequent modeling and fitting results. After obtaining the above data, the method calls a photovoltaic output mechanism model based on the rated parameters of photovoltaic components, array azimuth and inclination, and irradiance and ambient temperature data, calculates the clear theoretical output curve of each photovoltaic site in the future prediction time domain, and gives the theoretical power level from the physical mechanism perspective when not subject to grid constraints. At the same time, according to the current values of the distribution network topology, node voltage and branch power flow, the operation state of the voltage regulator and the load prediction data, the linearized power flow sensitivity model or the equivalent electrical distance model is used to calculate the allowable injected active power upper limit of each photovoltaic grid-connected point in the prediction time domain to meet the voltage limit constraint and the line thermal stability constraint, so that the voltage constraint and the thermal stability constraint of the grid side participate in the determination of the injected power upper limit in a quantitative form. At each prediction time, for each photovoltaic site, the clear theoretical output upper limit value obtained by the photovoltaic output mechanism model is compared with the injected active power upper limit value calculated by the distribution network constraint, and the smaller value of the two is taken as the power constraint upper limit of the site at that time, while zero power is taken as the power constraint lower limit, and the arithmetic mean of the upper and lower limits is taken as the baseline prediction power of the site at that time. Through repeated processing in the entire prediction time domain, the baseline power trajectory of the distributed photovoltaic cluster is obtained, which reflects the power level that each site can achieve in the future period under the joint action of photovoltaic output mechanism and grid safety constraints, and provides a reference curve with physical and grid significance for subsequent residual modeling. To depict the deviation between actual operation and baseline, the method obtains the residual value by subtracting the actual active power from the corresponding baseline predicted power for each photovoltaic station in multiple historical prediction periods, and normalizes the residual value according to the difference between the upper limit of the power constraint and the lower limit of the power constraint at the moment, forming the normalized residual time series of each station. The normalization process maps the deviation under different capacity stations and different constraint intervals to a unified dimension, facilitating unified modeling in spatial and temporal dimensions. When constructing the residual field sample, the change amount of the inverter operating state, the change rate of irradiance and environmental temperature, and the offset amount of the node voltage relative to the rated voltage are also used as exogenous influence features, which are combined with the normalized residual at the corresponding moment to reflect the influence of device operating state changes, meteorological conditions, and voltage fluctuations on the formation of the residual in the residual modeling process. In terms of spatial correlation modeling, the method determines the connection relationship between each photovoltaic grid-connected point according to the topology structure of the distribution network, calculates the electrical distance weight based on the electrical distance between nodes, and calculates the residual correlation weight based on the correlation coefficient between the normalized residual time series of different photovoltaic stations. Then, the electrical distance weight and the residual correlation weight are weighted and superimposed and normalized to construct a weighted adjacency matrix, so that each matrix element of the matrix represents the comprehensive weight between the corresponding two photovoltaic stations. The comprehensive weight reflects both the electrical connection distance and the residual evolution correlation degree. With this comprehensive weight as a coefficient, the propagation and coupling relationship between different station residuals can be described on the structure of the distribution network. On this basis, the method constructs a spatio-temporal recursive prediction model, taking the normalized residual of each photovoltaic station as the state quantity, the exogenous influence features as the input variable, and the weighted adjacency matrix as the coupling structure to describe the evolution process of different station residuals in the time dimension and the transmission relationship along the distribution network in the spatial dimension. This model is trained by minimizing the loss function of the error between the historical prediction residual and the corresponding actual normalized residual to obtain a spatio-temporal residual prediction model for predicting the evolution of the residual field. By taking the residual field as the modeling object, the model training focuses on the deviation part that is not explained by the photovoltaic output mechanism and the grid constraint, which facilitates the identification of additional change patterns caused by factors such as local shading, inverter state changes, and data fluctuations. In the process of ultra-short-term prediction, the method recalculates the baseline power trajectory in the future prediction time domain based on the multi-source data collected before the current time at the time to be predicted, selects a preset number of normalized residual time slices and their corresponding exogenous influence features in the historical prediction period, inputs them into the spatio-temporal residual prediction model to obtain the normalized residual prediction value of each prediction time, then performs inverse normalization on the normalized residual prediction value according to the difference between the upper and lower power constraints of each prediction time, and adds the inverse normalized residual prediction value to the baseline power trajectory to obtain the active power prediction value of each prediction step in the future prediction time domain of each photovoltaic station. For the prediction value exceeding the upper power constraint, it is truncated to the upper power constraint; for the prediction value lower than the lower power constraint, it is truncated to the lower power constraint, so that the final power prediction sequence is kept within the range defined by the aforementioned mechanism constraints and power grid operation constraints; Considering that different data sources and different stations have missing, abnormal and time alignment deviation in the actual collection process, the method calculates the data missing ratio, the abnormal value ratio and the timestamp offset in a sliding time window of a set length for each type of data source and each photovoltaic station, calculates the integrity index according to the missing ratio, calculates the effectiveness index according to the abnormal value ratio, calculates the synchronization index according to the timestamp offset, combines the three indexes to obtain the data quality index, and weights the samples of different data sources and different photovoltaic stations according to the data quality index when constructing the residual field sample and inputting the exogenous influence feature, so that the weight of the sample with low data quality is limited in the modeling and prediction process, and the influence of the sample on the residual prediction result is reduced. In the same sliding time window, the method also calculates the mean and variance of the residual prediction error of each photovoltaic station, wherein the residual prediction error is the difference between the actual normalized residual and the corresponding predicted normalized residual, and sets an error threshold corresponding to the data quality index for different photovoltaic stations and different data sources. When the mean or variance of the residual prediction error of a photovoltaic station or a type of data source is detected to exceed the corresponding error threshold, the latest residual field sample in the window is used to perform incremental update on the comprehensive weight related to the photovoltaic station in the weighted adjacency matrix and the parameters related to the photovoltaic station or the type of data source in the spatio-temporal residual prediction model, while the comprehensive weights related to other stations and the model parameters are not changed, so that the model adjusts the local prediction deviation while keeping the overall structure unchanged.

[0007] In an embodiment, the photovoltaic output mechanism model adopts an equivalent circuit structure comprising a photo-generated current source, a diode and a resistance branch, and parameters in the equivalent circuit are calibrated by open-circuit voltage, short-circuit current and temperature coefficient of the photovoltaic module, so that the equivalent circuit can reflect the output characteristics of the photovoltaic module under different irradiance and module temperature conditions. When the irradiance and module temperature are given, the voltage and current of the maximum power point on the direct current side are solved based on the calibrated equivalent circuit, and the direct current power corresponding to the maximum power point is converted into the theoretical output on the alternating current side in combination with the inverter efficiency curve. The above calculation is performed at each prediction time in the future prediction time domain to obtain a curve of the alternating current side theoretical output changing with time, which is taken as the emptying theoretical output curve. Through the above modeling method, the rated parameters and operating environment measurement data of the photovoltaic module can be converted into a power upper bound with physical constraint significance without relying on empirical coefficients, thereby providing a reference value consistent with the actual output capacity of the photovoltaic module for the construction of the baseline power trajectory.

[0008] In an embodiment, the linearized power flow sensitivity model considers the response relationship of node voltage and branch power flow to the active power injection of the photovoltaic grid-connected point. Specifically, based on the topological structure and electrical parameters of the distribution network, the sensitivity coefficients of each node voltage to the active power injection of each photovoltaic grid-connected point, and the sensitivity coefficients of each branch power flow to the active power injection of each photovoltaic grid-connected point are calculated, so that when the active power injection changes, the corresponding node voltage change and branch power flow change can be estimated in a linear approximation manner. In determining the upper bound of the active power injection, first, the allowed voltage deviation range of each node is determined according to the operating regulations and the voltage qualification range, and the allowed voltage deviation range of the corresponding node is converted into the upper bound of the active power injection based on the voltage constraint according to the sensitivity coefficients of the node voltage to the active power injection of the photovoltaic grid-connected point. At the same time, the allowed branch power flow change range is determined according to the thermal stability limit of each branch, and the allowed branch power flow change range is converted into the upper bound of the active power injection based on the branch power flow constraint according to the sensitivity coefficients of the branch power flow to the active power injection of the photovoltaic grid-connected point. At the same photovoltaic grid-connected point, the upper bound of the active power injection obtained from the node voltage constraint and the upper bound of the active power injection obtained from the branch power flow constraint are compared, and the smaller one is selected as the upper bound of the active power injection of the photovoltaic grid-connected point. The above calculation process is repeated at each prediction time in the future prediction time domain, so that the upper bound of the active power injection corresponding to the voltage out-of-limit constraint and the line thermal stability constraint is obtained at each prediction time, thereby providing a grid-side constraint basis for the construction of the subsequent power constraint upper bound and the baseline power trajectory.

[0009] In an embodiment, in order to simultaneously reflect the electrical connection relationship of the power distribution network and the consistency of the residual evolution of each photovoltaic site in the weighted adjacency matrix, weights are respectively constructed based on the electrical distance and the residual correlation, and the two are weighted combined; Specifically, first, the electrical distance between each photovoltaic grid-connected point is calculated according to the impedance parameters or equivalent electrical parameters of the power distribution network, the reciprocal of the electrical distance is taken, and normalization is performed between all pairs of grid-connected points to obtain the electrical distance weight, so that the grid-connected point pair with shorter electrical distance corresponds to larger electrical distance weight. Then, the residual correlation coefficient is calculated based on the normalized residual time series of different photovoltaic sites, and the residual correlation coefficient is scaled and normalized according to a preset proportion coefficient to obtain the residual correlation weight, so that the site with higher residual correlation corresponds to larger residual correlation weight; After obtaining the electrical distance weight and the residual correlation weight, the electrical distance weight coefficient and the residual correlation weight coefficient are set for each pair of photovoltaic grid-connected points, the electrical distance weight is multiplied by the corresponding electrical distance weight coefficient, the residual correlation weight is multiplied by the corresponding residual correlation weight coefficient, and the sum of the two is normalized in the whole network to obtain the comprehensive weight in the weighted adjacency matrix. The comprehensive weight obtained in the above manner decreases with the increase of the node electrical distance and increases with the increase of the normalized residual correlation coefficient, so that the weighted adjacency matrix can quantitatively reflect the correlation between the electrical structure of the power distribution network and the residual change between the photovoltaic sites when used in the spatiotemporal recursive prediction model, providing a parameter basis for setting the residual coupling strength in the model.

[0010] In an embodiment, in order to make the power deviation of different photovoltaic sites under different operating conditions comparable, the method normalizes the residual values of each site to construct a normalized residual time series. Specifically, at each prediction time, the residual value of each photovoltaic site at that time is divided by the difference between the upper and lower power constraints at the corresponding time to obtain a normalized residual value, and the normalized residual values are arranged in chronological order to form the normalized residual time series of the photovoltaic site. By taking the difference between the upper and lower power constraints as the normalization coefficient, the residual values obtained by different capacity sites under different constraint intervals can be uniformly mapped into a unified dimensionless value interval, facilitating the comparison and superposition of residual information of different photovoltaic sites in subsequent residual field modeling and spatial correlation analysis between multiple sites.

[0011] In an embodiment, in order to simultaneously depict the evolution law of the residual of a single photovoltaic site over time and the coupling relationship between the residuals of adjacent sites, the state update of the spatiotemporal recursive prediction model at each prediction step is divided into a local residual autoregressive term and an adjacent residual coupling term. Specifically, the local residual auto-regressive term is calculated based on the normalized residuals of the photovoltaic station in the preset number of time steps before the photovoltaic station and the exogenous influence features of the corresponding time steps, and is used to reflect the time evolution trend of the station under the action of its own historical residuals and external factors such as weather, voltage and inverter state changes; The adjacent residual coupling term is based on the comprehensive weight of the photovoltaic station in the weighted adjacent matrix, and the normalized residuals of each photovoltaic station adjacent to the photovoltaic station in the preset number of time steps before the photovoltaic station are weighted and summed, and are mapped through a pre-selected nonlinear function, and are used to reflect the influence of the adjacent station residuals on the station residuals along the power distribution network structure; at each prediction step, the local residual auto-regressive term and the adjacent residual coupling term are added to obtain the normalized residual prediction value of the next time step of the photovoltaic station, and the spatio-temporal recursive structure is used to quantitatively describe the time variation and spatial propagation process of the multi-station residual field based on the coupling relationship obtained by the historical residual sequence and the weighted adjacent matrix, and to provide residual prediction results for subsequent power prediction.

[0012] In another embodiment, the future prediction time domain is set to a time period of 5 minutes to 30 minutes, and the prediction time granularity is set to 1 minute or 5 minutes to adapt to the typical ultra-short-term prediction demand in the operation and scheduling scenario of the power distribution network. The calculation method for constructing the upper and lower bounds of the power constraint, the method for constructing the residual field sample, and the structure of the spatio-temporal residual prediction model remain unchanged when different prediction time granularities are used, and only the preset number of historical time steps in the normalized residual time sequence participating in the state update is adjusted according to the selected prediction time granularity. Through this setting, the same set of models can cover various ultra-short-term prediction demands and ensure that the power constraint construction, residual modeling and spatio-temporal recursive structure remain consistent under different prediction granularities without changing the overall structure and parameter form of the model.

[0013] In another embodiment, in order to quantitatively characterize the quality level of data collected by different data sources and different photovoltaic stations, the application calculates the integrity index, the effectiveness index and the synchronization index in the sliding statistical window respectively; Specifically, the integrity index is calculated according to the ratio of the number of effective sampling points to the number of theoretical sampling points in the statistical window, and is used to reflect the data missing condition; the effectiveness index is calculated according to the ratio of the number of data points judged as abnormal by the data checking rule to the number of effective sampling points in the statistical window, and is used to reflect the proportion of abnormal values; the synchronization index is calculated according to the ratio of the average value of the absolute value of the timestamp offset to the same time granularity in the statistical window, and is used to reflect the degree of time alignment deviation, and then the data quality index of the data source at the photovoltaic station is obtained by weighted sum of the integrity index, the effectiveness index and the synchronization index; The data quality indicators are used to limit the weight value range of the corresponding data source and the corresponding photovoltaic station in constructing the residual field sample and introducing the exogenous influence feature, so that the weight of data with low data quality in residual modeling and prediction is limited, thereby reducing the influence of missing data, abnormal data and timestamp offset on the residual field modeling and power prediction result.

[0014] In another embodiment, in order to monitor the prediction deviation of the spatio-temporal residual prediction model on different photovoltaic stations, the present application calculates the residual prediction error and its statistics of each photovoltaic station within a sliding time window; The residual prediction error is defined as the difference between the actual normalized residual and the corresponding predicted normalized residual, the residual prediction error mean is calculated as the arithmetic mean of all residual prediction errors within the sliding time window, the residual prediction error variance is calculated as the square mean of the difference between the residual prediction error and the residual prediction error mean, and the error threshold corresponding to each photovoltaic station and each type of data source is set in advance according to the rated capacity of the photovoltaic station and the data quality indicators related to the photovoltaic station; During operation, when the residual prediction error mean or the residual prediction error variance of a certain photovoltaic station exceeds the error threshold corresponding to the photovoltaic station and the related data source, local adjustment of the model parameters related to the photovoltaic station is triggered, and only the incremental update of the comprehensive weight related to the photovoltaic station in the weighted adjacency matrix and the parameters related to the photovoltaic station or the data source in the spatio-temporal residual prediction model is performed, while the comprehensive weights related to other photovoltaic stations and the model parameters remain unchanged; Through this threshold-triggered local parameter adjustment, the photovoltaic stations and data sources with large prediction deviation can be corrected specifically without changing the overall model structure and other regional parameters.

[0015] In another embodiment, the above incremental update is performed in a gradient update manner based on the residual field sample, specifically, a loss function related only to the target photovoltaic station and its adjacent photovoltaic stations is constructed based on the newly added residual field samples within the sliding time window, and the gradient calculation and parameter correction of the comprehensive weight and the model parameters related to the target photovoltaic station are performed, wherein only the residual field samples of the target photovoltaic station and its adjacent photovoltaic stations are used for calculation at each gradient update, and the residual field samples of other photovoltaic stations are not used; By limiting the sample range participating in the incremental update, the present application avoids affecting the residual coupling structure and model parameters of other regions in the whole network when performing local parameter adjustment, so that the model can gradually correct the local prediction deviation while maintaining the original overall convergence characteristics.

[0016] The beneficial effects of the present application are as follows: 1. The application simultaneously collects photovoltaic module side measurement, short-term weather forecast, and multi-source data such as distribution network node voltage, branch power flow, voltage regulating device state, and load prediction under the unified time dimension, gives the AC side empty theoretical output curve based on the photovoltaic output mechanism model, calculates the upper limit of injected active power that meets the voltage out-of-limit constraint and line thermal stability constraint using the linearized power flow sensitivity model or equivalent electrical distance model, constructs the baseline power trajectory between the upper limit of power constraint and the lower limit of power constraint, so that the subsequent prediction value is always within the range jointly defined by the photovoltaic side physical capacity and the distribution network safety constraint, reducing the situation that the numerical value seems reasonable but the corresponding operating point has local overvoltage or line overload in the high penetration rate feeder scene.

[0017] 2. The application forms the normalized residual time series of each photovoltaic station by the difference between the actual active power and the baseline power on the basis of constructing the baseline power trajectory, calculates the node electrical distance weight according to the distribution network topology and line parameters, calculates the residual correlation weight according to the historical normalized residual time series, and performs weighted superposition and normalization on the two types of weights to obtain the weighted adjacency matrix reflecting the comprehensive action of electrical distance and residual correlation, further constructs the space-time recursive prediction model taking the normalized residual as the state quantity, the exogenous influence feature as the input, and the weighted adjacency matrix as the coupling structure, and simultaneously utilizes the local residual autoregressive term and the adjacent residual coupling term to depict the evolution process of the residual in time and space in each prediction step, so that the spatial correlation information between multiple stations is more comprehensively utilized, and the accuracy and consistency of the distributed photovoltaic cluster ultra-short-term power prediction are improved.

[0018] 3. The application obtains the integrity index, validity index, and synchronization index by respectively counting the data missing ratio, abnormal value ratio, and time stamp offset of each data source and each photovoltaic station in the sliding time window, and constructs the data quality index accordingly, and when generating the residual field sample and arranging the exogenous influence feature, the different data sources and different stations are weighted according to the data quality index to suppress the influence of missing data, abnormal data, and time asynchronization on residual modeling and power prediction; meanwhile, the residual prediction error mean and variance of each photovoltaic station are calculated in the same sliding time window, the corresponding error threshold is set in combination with the rated capacity and the data quality index, when the prediction error of a station or a type of data source exceeds the threshold, only the weighted adjacency matrix comprehensive weight and the space-time residual prediction model parameters related to the station are executed incremental update based on the newly added residual field sample, so that the model can be adjusted in a targeted manner with the changes of data quality and working conditions in the long-term online operation process, and the stability and reliability of the ultra-short-term prediction result in actual application are improved. BRIEF DESCRIPTION OF DRAWINGS

[0019] Fig. 1 It is the core flow chart of the distributed photovoltaic power ultra-short-term prediction of the application. Fig. 2 The data quality processing and model incremental updating flowchart of the present application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0021] As shown in Figs. 1-2 In one specific embodiment, the distributed photovoltaic power ultra-short-term prediction method based on multi-source data fusion of the present application is applied to multiple distributed photovoltaic sites connected to a distribution network, for predicting the active power of each photovoltaic site in a future ultra-short-term period under the premise of meeting the operation constraints of the distribution network. The method operates under a unified time granularity and specifically includes the following steps: First, multi-source data is collected and aligned. For each prediction period, the following multi-source data is collected and aligned under a unified time granularity: historical active power measurement values of each distributed photovoltaic site, component-side current and voltage, inverter operation state and limit generation state; meteorological monitoring data such as irradiance, ambient temperature and wind speed near the photovoltaic site; short-term weather forecast data corresponding to the region; and node voltage, branch power flow, switch and voltage regulator operation state, and node load prediction data on the distribution network side. By aligning the above photovoltaic side, weather side and grid side data on the time axis, different data sources are made to correspond to the same prediction time point, which facilitates the subsequent construction of power constraints and residual fields under the same time reference; Secondly, the clear sky theoretical output curve is constructed based on the photovoltaic output mechanism model on the photovoltaic side. According to the rated parameters of the photovoltaic components of each photovoltaic site, the array azimuth angle and inclination angle, and the collected irradiance and ambient temperature data, the photovoltaic output mechanism model is called to calculate the clear sky theoretical output curve of each photovoltaic site in the future prediction time domain. The clear sky theoretical output curve reflects the theoretical power level of each photovoltaic site at different prediction times when it is not constrained by factors such as grid power limitation, providing a physical reference for subsequent determination of the upper bound of power constraints on the photovoltaic side; Then, the upper bound of injected active power is constructed based on the power flow model of the distribution network on the grid side. According to the current values of the topology structure, node voltage and branch flow of the distribution network, the operating state of the voltage regulating device and the load prediction data, the linearized power flow sensitivity model or the equivalent electrical distance model is used to calculate the upper bound of injected active power of each photovoltaic grid-connected point in the predicted time domain to meet the voltage out-of-limit constraint and the line thermal stability constraint. Through this step, the voltage qualified range and the line thermal stability limit of the distribution network are converted into the upper bound constraint of the injected active power of each grid-connected point, so that the photovoltaic power prediction is limited by the safe operation conditions of the grid; At each prediction time, for each photovoltaic station, the theoretical output value of the clear sky theoretical output curve at the time corresponding to the theoretical output value calculated by the grid side is compared with the upper bound of the injected active power, and the smaller value of the two is taken as the upper bound of the power constraint of the station at the time, and zero power is taken as the lower bound of the power constraint. The arithmetic mean of the upper and lower bounds is taken as the baseline prediction power of the photovoltaic station at the prediction time. The above calculation is repeated for each prediction time in the prediction time domain to obtain the baseline power trajectory of the distributed photovoltaic cluster in the entire prediction time domain. The baseline power trajectory reflects the achievable power level of each photovoltaic station in the future prediction period under the joint action of photovoltaic output capacity and grid safety constraints. In the aspect of residual modeling, in a plurality of historical prediction periods, for each photovoltaic station, the actual active power is subtracted from the baseline prediction power at the corresponding time to obtain a residual value, and the residual value is normalized according to the difference between the upper bound of the power constraint and the lower bound of the power constraint at the time to form a normalized residual time series of each photovoltaic station. The normalized residual time series is used to eliminate the differences in capacity and constraint intervals of different stations, so that the residuals are comparable. When constructing the residual field sample, the change amount of the inverter operating state, the change rate of the irradiance and the ambient temperature, and the offset amount of the node voltage relative to the rated voltage are taken as exogenous influence characteristics, together with the normalized residual at the corresponding time, to form a residual field sample. Thus, the influence of equipment state change, meteorological disturbance and voltage disturbance on the evolution of the residual is explicitly reflected in the residual field. In the aspect of spatial correlation modeling, the connection relationship between each photovoltaic grid-connected point is determined according to the topology structure of the distribution network, the electrical distance weight is calculated based on the electrical distance between nodes, and the residual correlation weight is calculated based on the correlation coefficient between the normalized residual time series of different photovoltaic stations. The electrical distance weight and the residual correlation weight are weighted and superimposed and normalized to construct a weighted adjacency matrix, so that each matrix element of the weighted adjacency matrix represents the comprehensive weight between the corresponding two photovoltaic stations. The comprehensive weight reflects the electrical connection distance and the correlation degree of residual evolution, and provides a quantitative basis for setting the residual coupling strength in the subsequent spatiotemporal prediction model; In the aspect of spatio-temporal prediction modeling, a spatio-temporal recursive prediction model is constructed, taking the normalized residual of each photovoltaic station as the state quantity, taking the exogenous influence feature as the input variable, and taking the weighted adjacency matrix as the coupling structure. The spatio-temporal recursive prediction model is trained by minimizing the loss function of the error between the model prediction residual and the actual normalized residual in multiple historical prediction periods, to obtain a spatio-temporal residual prediction model for predicting the evolution of the residual field. The model simultaneously utilizes the historical residual sequence of each station and the comprehensive weight in the weighted adjacency matrix to describe the change rule of the residual in the time dimension and the propagation relationship of the residual along the power distribution network structure in the space dimension. In actual ultra-short-term prediction, the baseline power trajectory in the future prediction time domain is recalculated according to the above method based on the multi-source data collected before the current time at the time to be predicted. At the same time, the normalized residual time slices and the corresponding exogenous influence features in a preset number of historical prediction periods are selected as inputs and fed into the spatio-temporal residual prediction model to obtain the normalized residual prediction value of each prediction time in the future. The normalized residual prediction value is processed by inverse normalization according to the difference between the upper and lower bounds of the power constraint corresponding to each prediction time. The inverse normalized residual prediction value is superimposed on the baseline power trajectory to obtain the active power prediction value of each photovoltaic station at each prediction step in the future prediction time domain. For the prediction value exceeding the upper bound of the power constraint, it is truncated to the upper bound of the power constraint. For the prediction value lower than the lower bound of the power constraint, it is truncated to the lower bound of the power constraint, so as to ensure that the final prediction result is within the range of the aforementioned photovoltaic output mechanism constraint and power distribution network operation constraint; In the aspect of data quality processing, for each type of data source and each photovoltaic station, the data missing ratio, the abnormal value ratio and the timestamp offset are calculated in a sliding time window of a set length. The integrity index is calculated according to the data missing ratio, the effectiveness index is calculated according to the abnormal value ratio, and the synchronization index is calculated according to the timestamp offset. The integrity index, the effectiveness index and the synchronization index are combined into a data quality index. When constructing the residual field sample and introducing the exogenous influence feature, the samples of different data sources and different photovoltaic stations are weighted according to the data quality index, so that the weight of the sample with low data quality in the model training and prediction is limited, thereby reducing the interference of missing, abnormal and time offset data on the residual field modeling and power prediction result. During the model running, in order to monitor the prediction deviation of the spatiotemporal residual prediction model at different photovoltaic sites, the mean and variance of the residual prediction error of each photovoltaic site are calculated in a sliding time window, and the error threshold is set in advance according to the rated capacity of each photovoltaic site and the data quality index of the corresponding data source. When the mean or variance of the residual prediction error of a photovoltaic site or a type of data source exceeds the corresponding error threshold, the latest residual field sample in the window is used to perform incremental update on the comprehensive weight related to the photovoltaic site in the weighted adjacency matrix and the parameters related to the photovoltaic site or the type of data source in the spatiotemporal residual prediction model, without changing the comprehensive weights related to other photovoltaic sites and the model parameters. Through this local incremental update mechanism triggered by the error threshold, the sites and data sources with larger prediction deviation can be corrected specifically while maintaining the stability of the overall structure of the model.

[0022] In an embodiment, the photovoltaic output mechanism model adopts an equivalent circuit structure including a photo-generated current source, a diode and a resistance branch, which is used to calculate the theoretical output power of the photovoltaic module side under different irradiance and component temperature conditions. The equivalent circuit model includes a photo-generated current source related to the photo-generated electrons of the photovoltaic module, a diode branch related to the recombination of the module junction region, and a series resistance branch and a parallel resistance branch for characterizing the conductor loss and the leakage channel. By setting the model parameters to be adjusted for each branch, the current-voltage characteristics of the photovoltaic module are described in a unified circuit framework. During parameter tuning, the open-circuit voltage, short-circuit current, maximum power point voltage and maximum power point current provided by the module manufacturer, as well as the temperature coefficient information of the open-circuit voltage and the short-circuit current, are used to solve and adjust the photo-generated current source parameters, diode reverse saturation current parameters, series resistance and parallel resistance parameters in the equivalent circuit model. Specifically, under standard test conditions, the module terminal voltage is equal to the open-circuit voltage, the module terminal current is equal to the short-circuit current, and the rated value of the module terminal voltage and current corresponding to the maximum power point. The above rated parameters are substituted into the current-voltage relationship of the equivalent circuit, and a set of model parameters that satisfy the open-circuit voltage, short-circuit current and maximum power point parameters within the preset tolerance range are determined by numerical iteration or least squares fitting. Those skilled in the art can use existing single-diode equivalent circuit models and general numerical calculation tools to implement this parameter tuning process without making additional assumptions about the equivalent circuit structure. During operation, in order to obtain the output characteristics of the component under different environmental conditions, the embodiment calculates the component temperature at the corresponding moment according to the collected irradiance and environmental temperature data. For example, the environmental temperature and irradiance can be converted into the component temperature by combining factors such as the component installation mode and the backplane heat exchange condition, using a linear approximation relationship or an empirical formula, and the temperature-related parameters in the photovoltaic current source and the diode branch are corrected accordingly. Under the given irradiance and component temperature conditions, the terminal voltage of the equivalent circuit is taken as a series of discrete candidate values, the current value corresponding to each candidate voltage point is solved according to the adjusted equivalent circuit model, and the output power is calculated. The voltage and current combination with the maximum output power is selected as the maximum power point of the direct current side under the environmental conditions. By performing the above solving process at each prediction time in the future prediction time domain, the variation sequence of the maximum power point power of the direct current side of the photovoltaic component under different irradiance and component temperature sequences can be obtained. Due to the difference in conversion efficiency of the inverter under different load rates, the embodiment introduces an inverter efficiency curve to convert the theoretical output power of the direct current side. The inverter efficiency curve can be obtained by interpolation or piecewise function fitting according to the efficiency-load curve data provided by the inverter manufacturer. At each prediction time, the power corresponding to the maximum power point of the direct current side at that time is substituted into the inverter efficiency curve to calculate the corresponding efficiency value, and the direct current power is multiplied by the efficiency to obtain the theoretical output of the alternating current side. Arranging the theoretical output of the alternating current side at each prediction time in the future prediction time domain in chronological order forms a curve of the theoretical output of the alternating current side varying with time, which is used as the emptying theoretical output curve in the method. Through the construction and parameter adjustment of the photovoltaic output mechanism model, the embodiment unifies the rated parameters of the photovoltaic component, the temperature coefficient, and the measured irradiance and environmental temperature data, as well as the inverter efficiency characteristics into a calculable equivalent circuit framework without introducing empirical reduction coefficients, so as to obtain the emptying theoretical output of the alternating current side corresponding to the environmental conditions at any time in the future prediction time domain. The emptying theoretical output curve serves as the upper bound of the physical capacity of the photovoltaic side, which is used together with the voltage limit constraint of the power distribution network and the line thermal stability constraint to construct the upper bound of the power constraint in the method, and is used in combination with the residual field modeling and the spatiotemporal residual prediction model to form the overall technical path of mechanism upper bound + grid constraint + residual spatiotemporal modeling. Compared with the conventional prediction method based on statistical fitting of historical output data or direct empirical coefficient correction of meteorological prediction, the embodiment explicitly introduces the component equivalent circuit mechanism and the inverter efficiency characteristics when constructing the emptying theoretical output curve, providing a reference curve with physical constraint significance for subsequent power constraint construction and residual prediction. Based on this, those skilled in the art can realize the technical features related to the photovoltaic output mechanism model in the claims.

[0023] In an embodiment, in order to embody the voltage excursion constraint and the line thermal stability constraint in a quantitative form into the active power injection allowable range of each photovoltaic grid-connected point at the power distribution network level, a linearized power flow sensitivity model is adopted at the power distribution network side, which is based on the power flow calculation result of the current operating point and uses a linear relationship to approximately describe the influence of the change of the active power injection of each photovoltaic grid-connected point on the node voltage and branch power flow in the vicinity of the operating point; specifically, based on the power distribution network topology structure and line parameters, first, the node voltage, branch power flow and other state quantities of the current operating point are obtained through conventional power distribution network power flow calculation, and the sensitivity coefficient matrix of the node voltage to the active power injection of each photovoltaic grid-connected point and the sensitivity coefficient matrix of the branch power flow to the active power injection of each photovoltaic grid-connected point are calculated at the operating point, so that when the active power injection changes slightly, the corresponding node voltage change and branch power flow change can be approximately calculated by using the sensitivity coefficient; In determining the upper bound of the active power injection under the voltage constraint, according to the power distribution network operation regulation and the voltage qualified range, the voltage allowable offset range of each key node is obtained, for example, the upper and lower offset limits of the node voltage are determined based on the rated voltage, for a certain photovoltaic grid-connected point, assuming that its active power injection increases in the positive direction, the sensitivity coefficient of the node voltage to the active power injection of the grid-connected point in the linearized power flow sensitivity model is used to approximately calculate the change of the voltage of each key node when the power of the grid-connected point increases by one unit; for each key node, according to its voltage allowable offset range and the corresponding sensitivity coefficient, the maximum increment of the active power injection of the grid-connected point is calculated under the premise of not exceeding the upper and lower limits of the node voltage, and the minimum value among the candidate increments of all key nodes is selected as the upper bound of the active power injection based on the voltage constraint, the above process is equivalent to constructing a linear inequality constraint about the active power injection increment under the given sensitivity coefficient condition, and solving the maximum feasible solution of the inequality constraint set, those skilled in the art can complete the calculation relying on the sensitivity analysis function provided by the existing power flow calculation software; In the determination of the upper limit of the injected active power under the branch thermal stability constraint, the branch power flow thermal stability limit is obtained according to the allowable current or allowable power of each key branch in the distribution network, under the assumption that the injected active power at the same photovoltaic grid-connected point is positively increased, the sensitivity coefficient of the branch power flow in the linearized power flow sensitivity model to the injected active power at the grid-connected point is used to approximately calculate the change of the branch power flow when the power at the grid-connected point is increased by one unit, and the current power flow value of each branch is combined with the corresponding thermal stability limit, for each key branch, the maximum increment of the injected active power at the grid-connected point is calculated under the condition that the thermal stability limit is not exceeded, and the minimum value is selected from the corresponding candidate increments of all key branches as the upper limit of the injected active power based on the branch power flow constraint, thereby without repeating the complete power flow simulation for each candidate injected power point, but by using the sensitivity coefficient, the branch thermal stability constraint is converted into a linear constraint on the injected active power increment, the calculation process is clear and easy to implement; In the comprehensive voltage constraint and branch power flow constraint, for each photovoltaic grid-connected point, the upper limit of the injected active power based on the node voltage constraint and the upper limit of the injected active power based on the branch power flow constraint are obtained respectively, the smaller value is selected from the two as the upper limit of the injected active power at the grid-connected point near the current operating point, combined with the current injected active power basis value of the grid-connected point, the injected active power upper limit curve of the grid-connected point can be constructed with the current operating point as the reference in the prediction time domain, which is used for the construction of the power constraint upper limit in the method, since the calculation of the sensitivity coefficient is based on the distribution network topology and line parameters, the voltage allowable offset range is determined according to the voltage qualification standard, and the branch power flow thermal stability limit is determined according to the equipment nameplate parameters or design value, the data used is clear in origin and conforms to the operating conditions of the actual power system; By using the above linearized power flow sensitivity model, the voltage over-limit constraint and the line thermal stability constraint can be simultaneously processed in a unified linear approximation framework on the distribution network side, and the constraint conditions acting on the node voltage and the branch power flow are converted into the upper limit values acting on the injected active power of each photovoltaic grid-connected point, so that the construction of the power constraint upper limit in the method has a clear physical basis on the power grid side, compared with the conventional method of setting a fixed injection upper limit according to the static allowable access capacity or only considering the voltage constraint without explicitly considering the branch thermal stability constraint, the present embodiment simultaneously uses the response relationship of the node voltage and the branch power flow to the injected active power in the same sensitivity model, and dynamically calculates the injected active power upper limit of each grid-connected point combined with the current operating point and the future load prediction data, thereby providing a quantitative input matching the actual operating state of the distribution network for the subsequent power constraint based on the baseline power trajectory and the residual prediction.

[0024] In an embodiment, in order to simultaneously reflect the power distribution network electrical structure and the consistency of the residual evolution of each photovoltaic station in the weighted adjacency matrix, an electrical distance weight and a residual correlation weight are respectively constructed based on the node electrical distance and the residual correlation, and the two are weighted and superimposed; Specifically, according to the topological structure and line impedance parameters of the power distribution network, the electrical distance between nodes where each photovoltaic grid-connected point is located is calculated. The electrical distance can be calculated according to the equivalent impedance, equivalent susceptance or pre-defined electrical distance index between nodes. A positive node electrical distance is obtained for each pair of photovoltaic grid-connected points. Then, the reciprocal of the node electrical distance is obtained to obtain an electrical distance reciprocal matrix. Then, the reciprocal of the electrical distance of each pair of grid-connected points is divided by the normalization coefficient, which is the sum of the reciprocal of the electrical distance of all pairs of grid-connected points, to obtain the electrical distance weight. The grid-connected point with closer electrical distance corresponds to larger electrical distance weight, and the grid-connected point with farther electrical distance corresponds to smaller electrical distance weight. In terms of residual correlation, based on the normalized residual time series of each photovoltaic station, the residual correlation coefficient between different photovoltaic stations is calculated. Pearson correlation coefficient or other linear correlation coefficient can be used as a measure of residual correlation coefficient. Within a preset historical time window, the normalized residual time series of any two photovoltaic stations is calculated to obtain a residual correlation coefficient matrix. Then, the residual correlation coefficient is scaled and normalized, for example, the negative correlation is truncated to zero, and the non-negative correlation is scaled linearly. The sum of the scaled rank correlation coefficients of all station pairs is used as the normalization coefficient to obtain the residual correlation weight. The station pair with higher residual correlation corresponds to larger residual correlation weight. After obtaining the electrical distance weight and the residual correlation weight, the electrical distance weight coefficient and the residual correlation weight coefficient are set for each pair of photovoltaic grid-connected points. According to the pre-set proportional relationship, the electrical distance weight is multiplied by the corresponding electrical distance weight coefficient, and the residual correlation weight is multiplied by the corresponding residual correlation weight coefficient. The sum of the two is normalized in the whole network to obtain the comprehensive weight in the weighted adjacency matrix. The comprehensive weight decreases with the increase of the node electrical distance and increases with the increase of the normalized residual correlation coefficient. The weighted adjacency matrix can reflect the electrical structure relationship of the power distribution network and the correlation relationship between the residual evolution of each photovoltaic station in the space-time recursive prediction model, providing a quantitative basis for setting the residual coupling strength.

[0025] In another embodiment, in order to make the power deviation of different photovoltaic stations in different power constraint intervals comparable, the residual value of each station is normalized to construct a normalized residual time series. Specifically, at each prediction time, for each photovoltaic station, a residual value is first obtained according to the difference between the actual active power and the baseline predicted power at the corresponding time, and then the residual value is divided by the difference between the upper limit of the power constraint and the lower limit of the power constraint at the time, to obtain a normalized residual value, the difference between the upper limit of the power constraint and the lower limit of the power constraint describes the range of allowable power change of the station at the time, and the residual value is normalized with respect to the range, so that the normalized residual value is between -1 and 1; the normalized residual values obtained at multiple prediction times for the same photovoltaic station are arranged in time sequence to form a normalized residual time series of the photovoltaic station, which is used for subsequent residual field modeling and spatial correlation analysis between multiple stations.

[0026] In yet another embodiment, in order to simultaneously describe the evolution law of the residual of a single photovoltaic station over time and the coupling relationship between the residuals of adjacent stations, a spatiotemporal recursive prediction model is used to predict the normalized residual field, and the state update of the spatiotemporal recursive prediction model at each prediction step includes a local residual autoregressive term and an adjacent residual coupling term; The local residual autoregressive term is calculated based on the normalized residuals of the photovoltaic station at a preset number of time steps in the past and the exogenous influence features at the corresponding time steps, and is used to reflect the time evolution trend of the station under the action of its own historical residual and external factors such as weather, voltage, and inverter state change; the exogenous influence features can include the change amount of the inverter operating state, the change rate of irradiance and ambient temperature, the offset amount of the node voltage relative to the rated voltage, etc. The adjacent residual coupling term is based on a weighted adjacency matrix, and the normalized residuals of each photovoltaic station adjacent to the photovoltaic station at a preset number of time steps in the past are weighted and summed, and then mapped through a pre-selected nonlinear function; the comprehensive weight related to the photovoltaic station in the weighted adjacency matrix is used in the weighted sum, and the nonlinear function can use a piecewise linear function or other continuous transformation functions commonly used in the art, to describe the nonlinear influence relationship of the residuals of adjacent stations on the residual of the station; At each prediction step, the local residual autoregressive term and the adjacent residual coupling term are added to obtain the normalized residual prediction value of the photovoltaic station at the next time step. Through the above structure, the spatiotemporal recursive prediction model simultaneously considers the autoregressive relationship in the time dimension and the residual coupling relationship in the spatial dimension in the same update formula, and by combining the electrical distance weight and the residual correlation weight contained in the weighted adjacency matrix, the time variation and spatial propagation process of the multi-station residual field can be described in a unified model.

[0027] In another embodiment, the future prediction time domain is set to a time period of 5 minutes to 30 minutes, and the prediction time granularity is set to 1 minute or 5 minutes to cover the ultra-short-term scheduling and control requirements commonly seen in power distribution network operation. When different prediction time granularities are used, the calculation method for constructing the upper and lower bounds of the power constraint, the method for constructing the residual field sample, and the structure of the spatiotemporal residual prediction model remain unchanged, and only the preset number of historical time steps of the normalized residual time series participating in state updating is adjusted according to the prediction time granularity. For example, when the prediction time granularity is 1 minute, the normalized residual samples of the last several minutes can be selected as the input of the local residual autoregressive term and the adjacent residual coupling term. When the prediction time granularity is 5 minutes, the number of historical time steps can be correspondingly reduced to cover a historical time span comparable to that when the granularity is 1 minute. In this way, without changing the overall structure of the model, the historical time step length is adjusted to adapt to prediction tasks of different time resolutions, so that the same set of spatiotemporal residual prediction models can be applied to different ultra-short-term prediction scenarios.

[0028] In another embodiment, in order to quantitatively characterize the quality level of data collected by different data sources and different photovoltaic sites, integrity indicators, validity indicators, and synchronization indicators are calculated respectively within a sliding statistical window, and a data quality indicator is constructed based on the three indicators. The integrity indicator is calculated according to the ratio of the number of effective sampling points to the number of theoretical sampling points within the statistical window, and is used to reflect the data missing situation. The number of effective sampling points is the number of data points that pass the basic format check and are not judged as abnormal. The number of theoretical sampling points is calculated according to the uniform time granularity and window length. The validity indicator is calculated according to the ratio of the number of data points judged as abnormal by the data check rule to the number of effective sampling points within the statistical window, and is used to reflect the proportion of abnormal values. The synchronization indicator is calculated according to the ratio of the average value of the absolute value of the timestamp offset to the uniform time granularity within the statistical window, and is used to reflect the degree of time alignment deviation. After obtaining the integrity indicator, the validity indicator, and the synchronization indicator, a weighted sum of the three indicators is obtained to obtain the data quality indicator of the corresponding data source at the corresponding photovoltaic site. The weight coefficient can be set offline or empirically according to the influence of different indicators on the prediction result. When constructing the residual field sample and introducing exogenous influence features, the data quality indicator is used to weight the samples of different data sources and different photovoltaic sites. The weight of the sample with a smaller data quality indicator in residual modeling and prediction is limited, thereby reducing the influence of missing data, abnormal data, and timestamp offset on residual field modeling and power prediction results.

[0029] In another embodiment, in order to monitor the prediction deviation of the spatio-temporal residual prediction model on different photovoltaic sites and make local adjustment to the sites with large prediction deviation, the residual prediction error and its statistics are calculated for each photovoltaic site within a sliding time window, and the preset error threshold is used to trigger incremental update of the related parameters; The residual prediction error is defined as the difference between the actual normalized residual and the corresponding predicted normalized residual. Within a given sliding time window, the residual prediction error mean is calculated as the arithmetic average of all residual prediction errors, and the residual prediction error variance is calculated as the square average of the difference between the residual prediction error and the residual prediction error mean. According to the rated capacity of each photovoltaic site and the data quality index related to the photovoltaic site, a corresponding error threshold is set for each photovoltaic site and each type of data source, reflecting the acceptable prediction error level under different capacity and data quality conditions. During the online operation of the model, when the residual prediction error mean or the residual prediction error variance of a certain photovoltaic site is detected to exceed the error threshold corresponding to the photovoltaic site and the related data source, local adjustment is triggered for the parameters related to the photovoltaic site, and only the comprehensive weight related to the photovoltaic site in the weighted adjacency matrix and the parameters related to the photovoltaic site or the data source of this type in the spatio-temporal residual prediction model are subjected to incremental update, while the comprehensive weights related to other photovoltaic sites and the model parameters remain unchanged. The incremental update can be performed in a gradient update manner based on the residual field samples. The target photovoltaic site for which the update is triggered is based on the newly added residual field samples within the sliding time window, and a loss function related only to the target photovoltaic site and its neighboring photovoltaic sites is constructed to perform gradient calculation and parameter correction on the comprehensive weight related to the target photovoltaic site and the model parameters. During each gradient update, only the residual field samples of the target photovoltaic site and its neighboring photovoltaic sites are used for calculation, and the residual field samples of other photovoltaic sites are not used, so as to avoid affecting the residual coupling structure and model parameters of other regions in the whole network during local parameter adjustment.

[0030] It should be noted that in this paper, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0031] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A method for ultra-short-term power prediction of distributed photovoltaic (PV) power based on multi-source data fusion, applied to multiple distributed PV sites connected to a distribution network, characterized in that... include: At a unified time granularity, multi-source data is collected and aligned for each forecast period. The multi-source data includes operational measurement data of each distributed photovoltaic site, meteorological monitoring and short-term weather forecast data of the area where the photovoltaic site is located, and voltage, power flow and load forecast data of the distribution network side. The theoretical output curves of each photovoltaic site are calculated using the photovoltaic output mechanism model based on the operational measurement data and meteorological data. Combined with the distribution network topology, voltage, power flow, voltage regulation device and load data, the upper limit of the injected active power of each photovoltaic grid connection point is calculated using the distribution network power flow constraint model. The power constraint interval is constructed with the upper limit of the injected active power and zero power to obtain the baseline power trajectory. Based on historical prediction cycles, for each photovoltaic site, the difference between the actual active power and the corresponding baseline power is calculated and normalized according to the power constraint interval width to form a normalized residual time series. Features reflecting inverter status, weather changes and node voltage deviation are extracted as exogenous influence features. The connection relationships between photovoltaic grid connection points are determined based on the distribution network topology, and a weighted adjacency matrix representing the coupling strength is constructed. A spatiotemporal recursive prediction model is constructed with normalized residuals as state variables, exogenous influence features as input variables, and the weighted adjacency matrix as the coupling structure. The spatiotemporal recursive prediction model is trained using historical normalized residuals and their exogenous influence features to obtain a spatiotemporal residual prediction model. Based on the current multi-source data, the baseline power trajectory is calculated. The historical normalized residuals and exogenous influence characteristics are input into the spatiotemporal residual prediction model to obtain the normalized residual prediction value for future time. The normalized residuals are then inversely normalized according to the width of the power constraint interval and superimposed on the baseline power trajectory to obtain the active power prediction value for each photovoltaic site. The prediction values ​​that exceed the power constraint interval are limited to the upper and lower bounds of the corresponding power constraint.

2. The method for ultra-short-term distributed photovoltaic power prediction based on multi-source data fusion according to claim 1, characterized in that: The photovoltaic output mechanism model is an equivalent circuit model that includes a photocurrent source, diodes, and resistor branches. The parameters in the equivalent circuit model are tuned using the open-circuit voltage, short-circuit current, and temperature coefficient of the photovoltaic module. Under given irradiance and module temperature conditions, the voltage and current at the maximum power point on the DC side are solved, and the power corresponding to the maximum power point is converted into the theoretical output on the AC side by combining the inverter efficiency curve. The change of the theoretical output on the AC side over time constitutes the cleared theoretical output curve.

3. The method for ultra-short-term distributed photovoltaic power prediction based on multi-source data fusion according to claim 2, characterized in that: The power flow constraint model of the distribution network is a linearized power flow sensitivity model, which includes the sensitivity coefficient of node voltage to the active power injected into each photovoltaic grid connection point and the sensitivity coefficient of branch power flow to the active power injected into each photovoltaic grid connection point. By multiplying the allowable offset range of node voltage by the corresponding voltage sensitivity coefficient, and multiplying the thermal stability limit of branch power flow by the corresponding power flow sensitivity coefficient, the upper bound of injected active power based on node voltage constraints and branch power flow constraints is obtained respectively. The smaller value is selected from the upper bound of injected active power as the upper bound of injected active power for the corresponding photovoltaic grid-connected point.

4. The method for ultra-short-term distributed photovoltaic power prediction based on multi-source data fusion according to claim 3, characterized in that: The construction of the weighted adjacency matrix includes: The electrical distance weight is obtained by normalizing the reciprocal of the electrical distance between nodes, and the residual correlation weight is obtained by scaling the correlation coefficient of the normalized residual time series of different photovoltaic sites. The electrical distance weight and the residual correlation weight are multiplied by their respective weight coefficients and added together. The sum is then normalized to obtain the comprehensive weight in the weighted adjacency matrix. The comprehensive weight decreases as the electrical distance between nodes increases and increases as the normalized residual correlation coefficient increases.

5. The method for ultra-short-term distributed photovoltaic power prediction based on multi-source data fusion according to claim 4, characterized in that: The normalized residual time series is formed as follows: for each photovoltaic site at each prediction time, the residual value is divided by the difference between the upper and lower bounds of the power constraint at the corresponding time to obtain a normalized residual value between -1 and 1, and arranged in chronological order to form the normalized residual time series of the photovoltaic site.

6. The method for ultra-short-term distributed photovoltaic power prediction based on multi-source data fusion according to claim 5, characterized in that: The state update of the spatiotemporal recursive prediction model at each prediction step includes a local residual autoregressive term and an adjacent residual coupling term. The local residual autoregressive term is calculated based on the normalized residual of the photovoltaic site and the exogenous influence characteristics of the corresponding time steps; The adjacency residual coupling term is based on the weighted adjacency matrix, which performs a weighted summation of the normalized residuals of each photovoltaic site adjacent to the photovoltaic site for a preset number of time steps, and maps them through a pre-selected nonlinear function. The local residual autoregressive term is added to the adjacent residual coupling term to obtain the normalized residual prediction value for the next time step of the photovoltaic site.

7. The method for ultra-short-term distributed photovoltaic power prediction based on multi-source data fusion according to claim 6, characterized in that: The future prediction time domain is a period of 5 to 30 minutes, and the prediction time granularity is 1 minute or 5 minutes. When different prediction time granularities are used, the construction methods of the upper and lower bounds of the power constraint, the construction methods of the residual field samples, and the structure of the spatiotemporal residual prediction model remain unchanged. Only the preset number of historical time steps in the normalized residual time series participating in the state update is adjusted according to the prediction time granularity.

8. The method for ultra-short-term distributed photovoltaic power prediction based on multi-source data fusion according to claim 7, characterized in that: The integrity index is calculated as the ratio of the number of valid sampling points to the theoretical number of sampling points within the statistical window. The validity index is calculated as the ratio of the number of data points judged as abnormal to the number of valid sampling points within the statistical window. The synchronicity index is calculated as the ratio of the average absolute value of the timestamp offset within the statistical window to the uniform time dimension. The data quality index is obtained by weighted summation of the integrity index, validity index, and synchronicity index. The data quality index limits the weight range of the corresponding data source and the corresponding photovoltaic site when weighting the residual field sample and the exogenous influence characteristics.

9. The method for ultra-short-term distributed photovoltaic power prediction based on multi-source data fusion according to claim 8, characterized in that: The residual prediction error is the difference between the actual normalized residual and the corresponding predicted normalized residual. The mean of the residual prediction error is calculated as the arithmetic mean of the residual prediction errors within the sliding time window. The variance of the residual prediction error is calculated as the squared average of the difference between the residual prediction error and the mean of the residual prediction error within the sliding time window. The error threshold is preset based on the rated capacity of the corresponding photovoltaic site and the data quality index of the corresponding data source. When the mean or variance of the residual prediction error of a certain photovoltaic site exceeds the error threshold corresponding to the photovoltaic site and related data source, only the comprehensive weights related to the photovoltaic site in the weighted adjacency matrix and the parameters related to the photovoltaic site or the data source in the spatiotemporal residual prediction model are incrementally updated.

10. A method for ultra-short-term distributed photovoltaic power prediction based on multi-source data fusion according to claim 9, characterized in that: The incremental update includes: based on the newly added residual field samples within the sliding time window, the comprehensive weights and model parameters related to the target photovoltaic site are corrected using a gradient update method, wherein each gradient update uses only the residual field samples of the photovoltaic site and its adjacent photovoltaic sites for calculation.

Images