Photovoltaic power minute-level rolling prediction method and system based on internal perception

Abstract

The invention relates to the technical field of new energy power generation operation control and intelligent prediction, in particular to a photovoltaic power minute-level rolling prediction method and system based on internal sensing, and the method comprises the steps: constructing a second-level multi-source real-time sensing network which comprises a multispectral light intensity sensor and a micro barometer in a photovoltaic power station; extracting features through an edge computing server and triggering internal primary early warning; an RNN hybrid prediction model (including delay perception modulation and an internal early warning enhancement mechanism) fusing the early warning features and external lagged meteorological data is established; starting minute-level rolling prediction in a severe convection potential period, dynamically adjusting the trust weight of a data source, outputting a power prediction value of 0-30 minutes in the future, and generating an emergency regulation and control instruction in case of power cliff type fall; according to the method, the problem of photovoltaic power prediction lag distortion in the short-time severe convection weather is solved, the perceptual perspectiveness and prediction reliability are improved, key early warning advance is gained for power grid dispatching and an energy storage system, and the operation safety of a power station and the friendliness of a power grid are enhanced.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation operation control and intelligent prediction technology, specifically a method and system for minute-level rolling prediction of photovoltaic power based on internal sensing. Background Technology

[0002] With the rapid growth of photovoltaic (PV) installed capacity, the volatility and uncertainty of PV output have placed higher demands on the safe operation and dispatch accuracy of the power grid. Rapid cloud movement, sudden radiation changes, and precipitous power drops caused by severe convective weather have become significant factors restricting the high-proportion grid-connected operation of PV. Therefore, accurately sensing weather changes and predicting power variations within minute or even shorter timescales is a critical technical problem that urgently needs to be solved in the field of PV forecasting. A search revealed a patent application (WO2014190651A1) that provides a PV power prediction method based on ground-based cloud maps. This method estimates cloud movement through cloud map sequences, thereby predicting changes in ground irradiance, and combines this with a photoelectric conversion model to achieve PV power prediction. It comprehensively utilizes digital image processing and multi-factor modeling techniques, improving the precision of PV power prediction to a certain extent and effectively supplementing traditional statistical prediction methods. It also exhibits good versatility and robustness. However, this method mainly relies on ground-based cloud maps and related meteorological data as core inputs, and the prediction effect is largely limited by the cloud map acquisition frequency, image processing latency, and cloud identification accuracy.

[0003] Traditional photovoltaic (PV) power forecasting methods generally focus on the analysis of external meteorological information or cloud image data, with insufficient utilization of real-time physical quantities within PV power plants. Furthermore, they fail to explicitly model the inherent time lag of external data, making it difficult to reflect local abrupt changes in scenarios with rapid development of severe convection. This can easily lead to minute-level rolling forecast results that are delayed or even distorted, hindering early response to the impact of severe convective weather. It is also difficult to accurately and timely perceive and predict local radiation abrupt changes and the "cliff-like" drop in PV power caused by rapidly moving cloud clusters. Consequently, it cannot provide effective early warning for grid dispatch and power plant self-regulation. Therefore, in response to the above situation, there is an urgent need to develop a minute-level rolling forecasting method and system for PV power based on internal perception to overcome the shortcomings in current practical applications. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for minute-level rolling prediction of photovoltaic power based on internal sensing, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for minute-level rolling forecasting of photovoltaic power based on internal sensing includes the following steps:

[0007] (1) Construct a second-level multi-source real-time sensing network in the photovoltaic array area of ​​the photovoltaic power station to collect data on total solar radiation intensity, module backsheet temperature, ambient temperature and humidity, multispectral irradiance and micro-pressure and aggregate them to the edge computing server.

[0008] (2) The edge computing server performs feature extraction on the collected data and generates radiation-related indicators, spectral distortion indicators and air pressure disturbance indicators. Based on the indicators, it triggers an internal primary warning signal for the approach of severe convective weather.

[0009] (3) Construct a hybrid prediction model that integrates internal real-time sensing features and external delayed meteorological data. The model includes a delayed sensing modulation mechanism and an internal early warning enhancement mechanism.

[0010] (4) When the internal primary warning signal is continuously triggered, start the minute-level rolling prediction process, update the internal warning status and external meteorological data delay information in real time, and input the hybrid prediction model;

[0011] (5) Output the photovoltaic power prediction sequence within the future preset time window through the hybrid prediction model, generate control instructions by combining the internal early warning level and send them to the power plant energy management system.

[0012] As a further aspect of the present invention: In step (1), the multi-source real-time sensing network adopts a gridded and functionally layered deployment strategy, with basic monitoring nodes arranged at equal intervals along the column and row directions of the components, and multispectral light intensity sensors and micro barometers arranged on the windward side of the photovoltaic array and in areas susceptible to cloud shadows.

[0013] As a further aspect of the present invention: in step (2), the radiation-related index is used to quantify the synchronous change trend of the whole station radiation, the spectral distortion index is used to characterize the change of solar radiation energy structure, and the pressure disturbance index is used to describe microscale pressure anomalies.

[0014] As a further aspect of the present invention: in step (3), the hybrid prediction model uses a recurrent neural network as its core framework, and the inputs include the internal primary warning signal intensity sequence, the historical sequence of total solar radiation intensity, and radar echo intensity and satellite cloud image data with delay markers.

[0015] As a further aspect of the present invention: in step (4), the minute-level rolling prediction process processes the second-level internal features through weighted time aggregation, so that the internal features are adapted to the minute-level rolling rhythm.

[0016] As a further aspect of the present invention: in step (5), the hybrid prediction model dynamically adjusts the dependence of internal sensing features on external meteorological data through a trust weight scheduling function, wherein the trust weight scheduling function is related to the intensity of internal early warning and the degree of delay of external data.

[0017] As a further aspect of the present invention: in step (5), when the prediction result indicates that a power precipitous drop will occur within a preset time in the future, an emergency control command with the highest priority is generated, wherein the preset time is 10 minutes.

[0018] A photovoltaic power minute-level rolling forecasting system based on internal sensing, used to implement the aforementioned photovoltaic power minute-level rolling forecasting method based on internal sensing, includes:

[0019] A multi-source real-time sensing network is deployed in the photovoltaic array area of ​​a photovoltaic power station to collect second-level multi-source sensing data.

[0020] An edge computing server is communicatively connected to the multi-source real-time sensing network and is used for feature extraction, triggering of internal primary early warning signals, and initiation of minute-level rolling prediction processes.

[0021] A hybrid prediction model, deployed on the edge computing server, is used to fuse internal real-time sensing features with external lagging meteorological data and output a photovoltaic power prediction sequence.

[0022] The power plant energy management system is communicatively connected to the edge computing server and is used to receive photovoltaic power prediction sequences and control instructions and execute corresponding operations.

[0023] As a further aspect of the present invention: the multi-source real-time sensing network includes basic monitoring nodes, multispectral light intensity sensors and micro barometers, wherein the basic monitoring nodes are used to collect data on total solar radiation intensity, component backplane temperature and ambient temperature and humidity.

[0024] As a further aspect of the present invention, the edge computing server is also used to perform timestamp correction on the data collected by the multi-source real-time sensing network and to assess the power drop risk of the photovoltaic power prediction sequence.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] This invention constructs a second-level multi-source real-time sensing network inside a photovoltaic power station, directly introducing physical quantities highly sensitive to severe convection, such as total solar radiation intensity, multispectral light intensity, and micro-pressure, into the prediction system. This makes photovoltaic power prediction no longer completely dependent on external meteorological data with inherent time delays, thereby improving the foresight and timeliness of sensing short-term severe convective weather from the source.

[0027] By constructing joint discrimination indicators such as radiation spatial consistency, rapid spectral decay and short-term air pressure fluctuations on the edge computing side, an internal primary warning can be triggered before the cloud cluster completely blocks the photovoltaic module, enabling early identification of the impact of strong convection and significantly shortening the perception lag time of power drop.

[0028] The established hybrid prediction model, which integrates internal real-time sensing features and external lagging meteorological data, explicitly models the delay characteristics of external meteorological data and introduces a delay sensing modulation mechanism and an internal early warning enhancement mechanism into the recurrent neural network. This enables the model to dynamically adjust its dependence on different information sources based on data timeliness and internal early warning intensity, effectively avoiding prediction distortion caused by external data lag under strong convection conditions.

[0029] By employing a minute-level rolling forecast process and a second-level internal feature time aggregation mechanism, the model input can be continuously refreshed during the rapid evolution phase of strong convection, ensuring that the forecast results always reflect the latest local meteorological conditions and improving the continuity and stability of photovoltaic power forecasts within a short time window.

[0030] The trust weight scheduling strategy based on the intensity of the warning and the degree of delay introduced in the forward calculation stage of the model enables the prediction results to stably extrapolate the power cliff trend in scenarios with strong internal warnings and lagging external data, which significantly improves the reliability of minute-level photovoltaic power prediction under extreme conditions.

[0031] By jointly evaluating minute-level power prediction sequences with internal early warning levels and directly connecting them to the power plant's energy management system, the system can automatically generate the highest-priority emergency control command when a sharp drop in photovoltaic power is predicted within the next ten minutes. This provides crucial advance warning for grid dispatch and energy storage system response, effectively improving the operational safety and grid friendliness of photovoltaic power plants. Attached Figure Description

[0032] Figure 1 This is a diagram showing the overall system architecture and data flow in an embodiment of the present invention.

[0033] Figure 2 This is a schematic diagram of the gridded and hierarchical deployment of monitoring nodes in an embodiment of the present invention.

[0034] Figure 3 This is a flowchart of edge feature extraction and early warning triggering in an embodiment of the present invention.

[0035] Figure 4 This is a diagram of the core structure of the hybrid prediction model in this embodiment of the invention.

[0036] Figure 5 This is a minute-level rolling prediction and control closed-loop sequence diagram in an embodiment of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0039] Please see Figures 1-5 The present invention provides a method and system for minute-level rolling forecasting of photovoltaic power based on internal sensing, which can solve the problem that existing photovoltaic power forecasting methods are difficult to accurately and timely predict the "cliff-like" drop in photovoltaic power under short-term severe convective weather due to the inherent time lag of relying on external meteorological data. This provides an effective early warning lead time for grid dispatching and power plant self-regulation.

[0040] In one embodiment of the present invention, the overall architecture of the internally sensed photovoltaic power minute-level rolling prediction system is as follows:

[0041] The photovoltaic power minute-level rolling prediction system of this invention mainly includes a multi-source real-time sensing network within the photovoltaic power plant area, an edge computing server, a hybrid prediction model, and a power plant energy management system. Each component interacts and transmits commands through an internal communication network. Figure 1 As shown in the diagram, the multi-source real-time sensing network is responsible for collecting multi-source sensing data at the second level, the edge computing server performs data processing, feature extraction and early warning triggering, the hybrid prediction model completes power prediction calculation, and the power plant energy management system receives the prediction results and control instructions and executes corresponding operations, forming a closed-loop operation mechanism of "sensing-prediction-control".

[0042] In one embodiment of the present invention, the present invention is based on a minute-level rolling prediction method for photovoltaic power based on internal sensing, and the specific implementation steps are as follows:

[0043] (I) Constructing a multi-source real-time sensing network

[0044] like Figure 2 As shown, a multi-source real-time sensing network is first constructed within the photovoltaic array area of ​​the photovoltaic power station. This network adopts a gridded and functionally layered deployment strategy, and the specific implementation is as follows:

[0045] Multiple basic monitoring nodes are distributed at equal intervals along the rows and columns of the photovoltaic array area. Each basic monitoring node synchronously collects data on the total solar radiation intensity, the temperature of the module backsheet, and the ambient temperature and humidity at its location. At the same time, several light intensity sensors and micro barometers with multispectral recognition capabilities are deployed on the windward side of the photovoltaic array and in key areas that are susceptible to cloud shadows. The multispectral light intensity sensors are used to capture changes in the solar spectrum caused by nearby clouds, and the micro barometers are used to sense small pressure fluctuations caused by the downdraft of clouds or the leading edge of gusts.

[0046] All monitoring nodes, multispectral light intensity sensors, and microbarometers are configured to acquire data at the second level. The collected sensor data is aggregated to a local edge computing server via an internal communication network. This internal communication network can employ common industrial communication network formats such as Ethernet or Wireless Local Area Network (WLAN) to ensure real-time and stable data transmission. After receiving data from each node, the edge computing server first performs unified timestamp correction to ensure the comparability of data from different nodes at the second level.

[0047] Based on the corrected, second-level sensing data, the edge computing server further calculates three core feature indicators to provide a data foundation for subsequent early warning and prediction:

[0048] The comprehensive index of radiation change within the station is used to quantify the synchronous change trend of radiation across the entire station over a very short time scale. Its expression is as follows:

[0049] ;

[0050] in, Indicates the first Each monitoring node at time Total solar radiation intensity collected. Indicates the number of effective monitoring nodes. This represents a second-level sampling time interval. This indicator is used to quantify the synchronous change trend of the whole station's radiation over a very short time scale. Its numerical mutations directly reflect the rapid intrusion process of cloud shadows.

[0051] The spectral distortion intensity index is used to characterize the influence of clouds on the structure of solar radiation energy, eliminating the influence of total radiation amplitude to highlight the instantaneous changes in spectral distribution. Its expression is:

[0052] ;

[0053] in, Indicates the first Each spectral band at time Irradiance, Indicates the number of bands. This represents the weighting coefficient for the sensitivity of the corresponding band to cloud shadows. This index, by eliminating the influence of the total radiation amplitude, highlights the instantaneous changes in the spectral distribution pattern and is used to identify the characteristics of strong convection-related cloud systems before the radiation has significantly decreased.

[0054] The local pressure disturbance energy index, used to describe microscale pressure anomalies, is expressed as follows:

[0055] ;

[0056] in, Indicates the first A microbarometer at time air pressure value, This represents the spatial average air pressure of all microbarometers at the same time. This indicates the number of microbarometers, and this index characterizes the non-uniform enhancement process of the pressure field through spatial variance.

[0057] This step, by directly deploying multi-source sensing devices inside the power plant, enables the second-level collection and real-time aggregation of physical quantities related to severe convective weather. Compared with traditional methods that rely on external meteorological data, this approach improves the foresight and timeliness of sensing short-term severe convective weather from the data source, providing high spatial resolution and high timeliness of basic data support for subsequent early warning and accurate forecasting.

[0058] (II) Feature Extraction and Internal Primary Early Warning Trigger

[0059] After receiving the high-frequency feature stream (on a second-level scale) uploaded from the multi-source real-time sensing network, the edge computing server initiates the feature extraction and preliminary early warning process, such as... Figure 3 As shown:

[0060] First, based on the total solar radiation intensity data of all monitoring nodes at the entire station, the spatial average rate of change and the coefficient of variation of the total solar radiation intensity of the entire station are calculated. A joint metric for spatial consistency and dispersion of radiation is constructed to distinguish between rapid overall shading and local random disturbances. Its calculation form is as follows:

[0061] ;

[0062] in, and Representing time respectively The spatial mean and standard deviation of the rate of change of total solar radiation intensity across the entire station show a synchronous amplification characteristic when the cloud front intrudes as a whole, thus amplifying the spatial synergistic changes related to strong convection.

[0063] Secondly, based on the data collected by the multispectral light intensity sensor, the spectral distortion intensity index S calculated in step (i) is... tA time gradient enhancement mechanism is introduced to construct a spectral attenuation enhancement index, which is used to capture the continuous attenuation process of energy in a specific band over a very short time. Its expression is:

[0064] ;

[0065] Where t represents the length of the time gradient calculation window, this index suppresses disordered fluctuations caused by high-frequency noise by accumulating only the band changes in irradiance that decrease monotonically, so that the rapid spectral decay caused by cloud shadow can be stably identified within a few seconds.

[0066] Simultaneously, for the second-level air pressure sequence collected by the microbarometer, the short-time fluctuation variance is calculated, and an air pressure fluctuation index is constructed, which is defined as:

[0067] ;

[0068] in, Indicates the first A microbarometer with a length of The mean value within the sliding time window is used to characterize the abnormal strengthening trend of air pressure over a short time scale through the form of time variance; L represents the length of the sliding time window, which can reflect the intensity of the disturbance of the local air pressure field caused by convective activity.

[0069] The edge computing server compares the aforementioned joint metric of radiation spatial consistency and dispersion, spectral attenuation enhancement, and air pressure fluctuation with their respective preset empirical thresholds in parallel. The preset empirical thresholds can be calibrated and optimized based on the climate characteristics, terrain conditions, and historical operating data of the area where the photovoltaic power station is located.

[0070] When any of the above-mentioned characteristic indicators exceeds its corresponding preset experience threshold, the edge computing server immediately triggers an internal primary warning signal for approaching severe convective weather, records the trigger time and duration of the warning signal, and outputs the warning intensity level.

[0071] This step, by constructing multi-dimensional joint discrimination indicators at the edge, achieves independent early warning without relying on external meteorological data. It can identify signals of approaching severe convective weather in advance before clouds completely block photovoltaic modules and before photovoltaic power shows a significant decline, significantly shortening the perception lag time of power drop and gaining valuable response time for subsequent minute-level rolling forecasts.

[0072] (III) Construction of Hybrid Prediction Model

[0073] A hybrid prediction model integrating internal real-time sensing features and external lagged meteorological data is constructed. This model uses a recurrent neural network (RNN) as its core framework, such as... Figure 4 As shown, the model construction and training process is as follows:

[0074] Model input design: The model input includes three types of data: the internal primary warning signal intensity sequence after time alignment, the historical sequence of total solar radiation intensity, and radar echo intensity and satellite cloud image data from external meteorological services. The external meteorological data needs to be labeled with its inherent delay time.

[0075] To achieve unified modeling of internal real-time features and external lagging data, an explicit time alignment and delay labeling mechanism is introduced on the model input side. The internal primary warning signal intensity sequence and the historical sequence of total solar radiation are used as instantaneous feature streams. At the same time, radar echo intensity and satellite cloud image data are encoded according to the delay between their actual acquisition time and the current prediction time to form an external feature vector with delay weight.

[0076] Delay-aware modulation mechanism: To quantify the impact of the timeliness of external data on prediction reliability, a delay-aware modulation factor is introduced into the input gating structure of the recurrent neural network, defined as follows:

[0077] ;

[0078] in, Indicates the time of external meteorological data The corresponding actual delay time, This represents the latency sensitivity coefficient, which is used to continuously attenuate the influence of lagged data within the model, so that external information automatically reduces its dominant role in the update of hidden states as latency increases.

[0079] Internal early warning enhancement mechanism: To highlight the dominant value of internal primary early warning signals in the approaching stage of severe convection, the intensity sequence of internal primary early warning signals is embedded into the state update process of a recurrent neural network to construct an internal feature enhancement term, the expression of which is:

[0080] ;

[0081] in, Indicates time The hidden state, This represents an input vector that includes radiation history and external meteorological characteristics. Indicates the strength of the internal primary warning signal. Represents the early warning mapping matrix. This represents the enhancement coefficient. This structure enables the model to quickly adjust the direction of the hidden state evolution when the warning signal is enhanced, thus strengthening its sensitivity to power abrupt changes.

[0082] Multi-step joint output mechanism: In the output layer, the model jointly predicts the power of multiple future time steps based on the current hidden state. To ensure the continuity and stability of the prediction results in minute-level rolling scenarios, a multi-step consistency constraint term is introduced, which takes the form:

[0083] ;

[0084] in, Indicates the future Predicted power value for each minute The multi-step mapping function representing shared parameters enables the model to output a complete short-term power evolution trajectory in a single forward calculation, namely the total power prediction of the photovoltaic power plant at five-minute intervals within the next 0 to 30 minutes.

[0085] Model training: The hybrid prediction model is trained using historical operating data of photovoltaic power plants. The training data includes historical second-level multi-source sensing data, historical internal early warning signal sequences, historical external meteorological data with delay markers, and corresponding actual photovoltaic power output data. The mean square error loss function between the predicted power and the actual power is minimized through the backpropagation algorithm to complete the optimization and update of the model parameters.

[0086] The hybrid prediction model constructed in this step effectively solves the problem of fusion modeling of internal real-time features and external lagging meteorological data by introducing a delay-aware modulation mechanism and an internal early warning enhancement mechanism. This enables the model to dynamically adjust its dependence on different data sources according to the timeliness of the data and the intensity of the early warning. It fundamentally makes up for the defect of traditional prediction models that are prone to prediction distortion due to the lag of external data under severe convective weather, and provides a stable and interpretable model foundation for subsequent minute-level rolling predictions.

[0087] (iv) Initiation of minute-level rolling forecast process

[0088] When the edge computing server triggers an internal primary early warning signal and enters a continuous state, it is determined to be a period of potential severe convective weather. At this time, a minute-level rolling forecast process is automatically initiated, such as... Figure 5 As shown:

[0089] Real-time data updates: At each rolling moment, the system prioritizes receiving the latest second-level data uploaded by the multi-source real-time sensing network, including total solar radiation intensity, component backplane temperature, ambient temperature and humidity, multispectral irradiance, and micro-pressure data. Based on the joint discrimination rules defined in step (II), the system instantly updates the strength and status of the internal primary warning signal, ensuring that internal characteristics always reflect the latest situation of cloud shadows and pressure disturbances. Simultaneously, it receives the latest arriving external meteorological data (radar echo intensity and satellite cloud image data), and calculates the actual delay by reading its data timestamp. The expression for the actual delay is:

[0090] ;

[0091] in, Indicates the trigger time of the current rolling prediction. The delay amount represents the observation time of external meteorological data. This delay is passed to the hybrid prediction model along with the radar echo intensity and satellite cloud image characteristics as an explicit label to maintain the effectiveness of the delay-aware modulation mechanism in step three.

[0092] Time-scale aggregation: To avoid scale mismatch between second-level internal features and minute-level rolling rhythm, the edge computing server performs weighted time aggregation on internal early warning signals and radiation-related features within the rolling window to form a representative description of the current minute state, defined as:

[0093] ;

[0094] in, Representing history The corresponding internal primary warning signal strength per second. Indicates the length of the scrolling aggregation window. This represents the time decay coefficient, which ensures that the model input remains sensitive to sudden changes by emphasizing the most recent warning information;

[0095] Model input construction: The internal feature vector after time aggregation, the historical sequence of total solar radiation intensity, and the external meteorological features with delay labels are spliced ​​together according to a unified temporal structure to form the complete model input vector at the current rolling moment, which is immediately sent to the hybrid prediction model constructed in step (III) to perform forward inference.

[0096] This step enables real-time tracking of the evolution of severe convective weather by initiating a minute-level rolling forecast process. Through real-time data updates and time-scale aggregation, it ensures the timeliness of the model input data and solves the adaptation problem of data at different time scales. This allows the hybrid forecast model to continuously receive input information reflecting the latest local meteorological conditions, significantly improving the continuity and dynamic response capability of power forecasts within a short time window.

[0097] (v) Forward computation of the hybrid prediction model

[0098] After receiving the complete input vector constructed in step (iv), the hybrid prediction model immediately performs forward computation to generate a photovoltaic power prediction sequence for the next 0 to 30 minutes at 5-minute intervals. The core process is as follows:

[0099] Trust-based weighted scheduling: Before updating the hidden state, the model constructs a unified trust-based scheduling function, which transforms the strength of the internal real-time early warning signal and the latency of the external meteorological data into calculable weight adjustment factors to characterize the relative dominance of the two types of data at the current rolling moment. Its definition is:

[0100] ;

[0101] in, This indicates the intensity of the minute-level internal early warning generated in step four. This indicates the actual delay time of external meteorological data. This represents the delay penalty coefficient. This function automatically increases the weight of internal features when internal warnings are enhanced or external data lag is aggravated, thus making the model more focused on the early mutation information captured by the local perception network.

[0102] Hidden State Fusion Update: Based on Trust Weight ω t The model performs weighted mapping on internal and external features separately, and completes the fusion during the hidden state evolution process. Its state update form is as follows:

[0103] ;

[0104] in, This represents a feature vector composed of internal early warning signals and radiation history. This represents an external meteorological feature subvector with a delay label. This structure ensures that the hidden state has a rapid response capability to local anomalies in the early stages of strong convection, while gradually regressing to utilize large-scale evolution trends as the timeliness of external data is restored.

[0105] Incremental multi-step output: The model employs a unified multi-step decoding strategy to continuously infer the power at multiple future time points, and its output expression is as follows:

[0106] ;

[0107] in, This represents the measured power at the current moment. Indicates minute-level step size. This represents a power increment prediction function based on hidden states. This form ensures the physical continuity of the prediction sequence in time by gradually accumulating the power change.

[0108] This step, through trust weight scheduling and weighted state update mechanism, enables the hybrid prediction model to stabilize the extrapolation of the precipitous decline in photovoltaic power in extreme scenarios where internal early warning is strong and external data is lagging. This effectively improves the reliability and accuracy of minute-level power prediction under severe convective weather, providing highly reliable predictive input for subsequent regulation and control decisions.

[0109] (vi) Output of prediction results and generation of control instructions

[0110] After obtaining the complete minute-level power prediction sequence output by the hybrid prediction model, the edge computing server jointly evaluates the prediction results with the current internal early warning level, and publishes the results in a structured manner to the power plant energy management system according to the power plant operation control interface specification. The specific process is as follows:

[0111] Power sag risk assessment: The system first performs a rate of change analysis on the future forecast sequence to construct a short-term power sag intensity index, which is used to quantify the maximum relative power drop that may occur within the next ten minutes. Its definition is:

[0112] ;

[0113] in, This indicates the current measured power. Indicates the future The predicted power corresponding to each minute is used to quantify the maximum relative power drop that may occur within the next ten minutes, thereby transforming the continuous prediction sequence into a risk measure that can be directly used for control judgment.

[0114] Control command triggered: Power drop intensity index F t Compared with the current minute-level internal early warning intensity A t The variables are integrated to form a unified control trigger discrimination variable, the expression of which is:

[0115] ;

[0116] in, This indicates the current minute-level internal warning intensity. This represents the warning amplification factor. This discriminant amplifies the risk of power drop when a strong convection warning is enhanced, making the control logic highly sensitive to sudden power cliffs.

[0117] when When the preset emergency control threshold is exceeded, the system immediately generates an emergency control command with the highest priority. The command carries key parameters such as the predicted power trajectory, the expected drop magnitude and time location, which are used to guide the energy storage system to switch charging and discharging in advance and reserve power balance space for grid dispatch.

[0118] Uncertainty Assessment and Feedback: To ensure the temporal continuity and traceability of regulatory actions, the system quantitatively assesses the overall uncertainty of the predicted sequence, which is defined as follows:

[0119] ;

[0120] in, Indicates the minute-level prediction step size. This indicates the number of prediction steps (i.e., 6 steps). This indicator changes synchronously with the steepness of the prediction curve and is used to mark the risk level of the prediction results in the energy management system.

[0121] Ultimately, minute-level power prediction sequences, internal early warning levels, drop risk indicators, and control instructions are uniformly sent to the power plant control layer and used as feedback input for the next round of rolling predictions and model effect correction, forming a closed-loop operation.

[0122] This step achieves seamless integration of forecasting and control by jointly evaluating the forecast results and the early warning level. When a sharp drop in power is predicted within the next ten minutes, emergency control instructions can be automatically generated, providing the grid dispatch and energy storage system with a crucial lead time and effectively improving the operational safety and grid friendliness of photovoltaic power plants under severe convective weather.

[0123] In summary, this invention forms a complete "sensing-prediction-control" closed loop by constructing a second-level multi-source real-time sensing network inside the power plant, triggering internal primary early warning, building a hybrid prediction model that integrates internal and external data, initiating a minute-level rolling prediction process, and generating control instructions. It effectively solves the problems of minute-level prediction lag and distortion of photovoltaic power under short-term severe convective weather in existing technologies, and has significant practicality and promotion value.

[0124] It should be noted that, in this invention, although the specification describes the embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for minute-level rolling prediction of photovoltaic power based on internal sensing, characterized in that, Includes the following steps: (1) Construct a second-level multi-source real-time sensing network in the photovoltaic array area of ​​the photovoltaic power station to collect data on total solar radiation intensity, module backsheet temperature, ambient temperature and humidity, multispectral irradiance and micro-pressure and aggregate them to the edge computing server. (2) The edge computing server performs feature extraction on the collected data and generates radiation-related indicators, spectral distortion indicators and air pressure disturbance indicators. Based on the indicators, it triggers an internal primary warning signal for the approach of severe convective weather. (3) Construct a hybrid prediction model that integrates internal real-time sensing features and external delayed meteorological data. The model includes a delayed sensing modulation mechanism and an internal early warning enhancement mechanism. (4) When the internal primary warning signal is continuously triggered, start the minute-level rolling prediction process, update the internal warning status and external meteorological data delay information in real time, and input the hybrid prediction model; (5) Output the photovoltaic power prediction sequence within the future preset time window through the hybrid prediction model, generate control instructions by combining the internal early warning level and send them to the power plant energy management system.

2. The photovoltaic power minute-level rolling prediction method based on internal sensing according to claim 1, characterized in that, In step (1), the multi-source real-time sensing network adopts a gridded and functionally layered deployment strategy. Basic monitoring nodes are deployed at equal intervals along the row and column directions of the components, and multispectral light intensity sensors and micro barometers are deployed on the windward side of the photovoltaic array and in areas susceptible to cloud shadows.

3. The method for minute-level rolling forecasting of photovoltaic power based on internal sensing according to claim 1, characterized in that, In step (2), the radiation-related index is used to quantify the synchronous change trend of the total station radiation, the spectral distortion index is used to characterize the change of solar radiation energy structure, and the pressure disturbance index is used to describe microscale pressure anomalies.

4. The photovoltaic power minute-level rolling prediction method based on internal sensing according to claim 1, characterized in that, In step (3), the hybrid prediction model uses a recurrent neural network as its core framework, and the inputs include the internal primary warning signal intensity sequence, the historical sequence of total solar radiation intensity, and radar echo intensity and satellite cloud image data with delay markers.

5. The method for minute-level rolling forecasting of photovoltaic power based on internal sensing according to claim 1, characterized in that, In step (4), the minute-level rolling prediction process processes the second-level internal features through weighted time aggregation, so that the internal features are adapted to the minute-level rolling rhythm.

6. The method for minute-level rolling forecasting of photovoltaic power based on internal sensing according to claim 1, characterized in that, In step (5), the hybrid prediction model dynamically adjusts the dependence of internal sensing features on external meteorological data through a trust weight scheduling function, which is related to the intensity of internal early warning and the degree of delay of external data.

7. The method for minute-level rolling forecasting of photovoltaic power based on internal sensing according to claim 1, characterized in that, In step (5), when the prediction result indicates that a power precipitous drop will occur within a preset time in the future, an emergency control command with the highest priority is generated, and the preset time is 10 minutes.

8. A photovoltaic power minute-level rolling forecasting system based on internal sensing, characterized in that, To implement the internal sensing-based minute-level rolling forecasting method for photovoltaic power as described in any one of claims 1-7, comprising: A multi-source real-time sensing network is deployed in the photovoltaic array area of ​​a photovoltaic power station to collect second-level multi-source sensing data. An edge computing server is communicatively connected to the multi-source real-time sensing network and is used for feature extraction, triggering of internal primary early warning signals, and initiation of minute-level rolling prediction processes. A hybrid prediction model, deployed on the edge computing server, is used to fuse internal real-time sensing features with external lagging meteorological data and output a photovoltaic power prediction sequence. The power plant energy management system is communicatively connected to the edge computing server and is used to receive photovoltaic power prediction sequences and control instructions and execute corresponding operations.

9. The photovoltaic power minute-level rolling forecasting system based on internal sensing according to claim 8, characterized in that, The multi-source real-time sensing network includes basic monitoring nodes, multispectral light intensity sensors, and micro barometers. The basic monitoring nodes are used to collect data on total solar radiation intensity, component backplane temperature, and ambient temperature and humidity.

10. The photovoltaic power minute-level rolling forecasting system based on internal sensing according to claim 8, characterized in that, The edge computing server is also used to perform timestamp correction on the data collected by the multi-source real-time sensing network and to assess the power drop risk of the photovoltaic power prediction sequence.

Images