A photovoltaic module and system

Abstract

The present application relates to the technical field of control adjustment system, and more particularly to a photovoltaic module and system, comprising a sensor group, a dynamic feature extraction module, a prediction model module, a collaborative optimization control module and an instruction execution module. The sensor group collects the front irradiance, back reflectivity and temperature gradient data of the double-sided photovoltaic array, synchronously obtains the state of charge and charge / discharge rate of the energy storage system, and the instruction execution module drives the equipment to act and calibrates the noise filtering threshold through the feedback channel. The system level integrates the edge computing node, the cloud scheduling center and the redundant communication network, realizes the local real-time control and the global resource optimization, combines the virtual sensing of the fault processing module, the off-grid strategy and the case library self-learning mechanism, solves the multi-time scale collaborative control problem of the dynamic fluctuation of the double-sided photovoltaic output and the response lag of the energy storage, and improves the energy utilization rate and the grid frequency modulation accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of control and regulation systems, and in particular to a photovoltaic module and system. BACKGROUND

[0002] The existing photovoltaic module and system adopts high-efficiency N-type monocrystalline silicon technology, applies super-multi-main-grid half-piece structure design, optimizes the current transmission path by increasing the number of main grids, reduces the series resistance and grid line shading loss, combines the half-cell cell splitting process to reduce the shadow shading effect, thereby improving the conversion efficiency of the module and reducing the risk of hot spot effect. The temperature coefficient optimization further enhances the power generation stability of the module in high temperature environment, reduces the influence of light-induced degradation on long-term power generation. The system level adopts modular design, taking 630Wp double-sided double-glass module as the basic unit, forming a string loop by connecting each 24 modules in series, combining 300kW string inverter and 4500kVA box-type step-up transformer to realize power conversion and step-up. The double-sided power generation characteristics make the module can utilize the direct light on the front side and the water surface reflected light on the back side at the same time, combined with the water cooling effect in the fish-light complementary mode, the system power generation efficiency is improved by about 10%-15%.

[0003] The existing technology has not fully solved the coordination and optimization problem between the dynamic fluctuation of photovoltaic array power and the response delay of energy storage system in the control and regulation system, which is specifically manifested as the nonlinear characteristics of photovoltaic output caused by the difference between front and back irradiance, the dynamic change of water reflectivity and the temperature fluctuation in the fish-light complementary environment in the double-sided power generation mode, and the multi-time scale coupling relationship between the charge and discharge rate of the rental energy storage system, the state of charge (SOC) constraint and the grid dispatching instruction. The existing control strategy based on fixed threshold or single feedback cannot realize the dynamic matching of power generation and energy storage in the time dimension of seconds to hours, resulting in the coordinated control bottleneck of decreased regulation accuracy, increased energy loss and insufficient frequency modulation capacity utilization when the system is responding to irradiation mutation, load steep increase or frequency emergency. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a photovoltaic module and system for solving the coordinated control problem between the multi-time scale dynamic fluctuation of double-sided photovoltaic array power and the response characteristics of rental energy storage system.

[0005] To solve the above technical problems, the specific technical solutions of the present application are as follows:

[0006] In a first aspect, the present application provides a photovoltaic module, comprising:

[0007] a sensor group for collecting the front irradiance, back reflectance and ambient temperature data of the bifacial photovoltaic array, and synchronously obtaining the state of charge and charging / discharging rate of the energy storage system;

[0008] a dynamic feature extraction module connected to the sensor group, for time stamp alignment and noise filtering of the collected data, generating a dynamic feature vector including the irradiance difference, temperature fluctuation and energy storage state parameters, and outputting the dynamic feature vector to the prediction model module;

[0009] a prediction model module receiving the front irradiance, back reflectance and temperature fluctuation data in the dynamic feature vector, modeling the second-level irradiance change of the front irradiance using a preset long short-term memory network, generating second-level power generation deviation prediction data, simultaneously inputting the temperature fluctuation data into a preset Kalman filter for component temperature rise state evaluation, outputting a temperature rise state correction value, and updating the preset energy storage efficiency decay model parameters according to the state of charge and charging / discharging rate data, to generate energy storage efficiency decay prediction results;

[0010] a cooperative optimization control module connected to the prediction model module, including a second-level control loop, a minute-level control loop and a hour-level control loop, inputting the second-level power generation deviation prediction data into the preset second-level control loop, generating an inverter power correction instruction through a preset adaptive droop control algorithm, inputting the temperature rise state correction value and energy storage efficiency decay prediction results into the minute-level control loop, combining a preset rolling horizon optimization algorithm to generate an energy storage charging / discharging power distribution table, and inputting a grid frequency modulation time period constraint into the hour-level control loop, generating a cluster frequency modulation capacity distribution proportion table through a preset mixed integer linear programming;

[0011] an instruction execution module receiving the inverter power correction instruction, energy storage charging / discharging power distribution table and cluster frequency modulation capacity distribution proportion table, adjusting the maximum power point tracking voltage parameters of the bifacial photovoltaic array and the charging / discharging rate of the energy storage system, and transmitting the adjusted parameters of the bifacial photovoltaic array and the actual output power data to the dynamic feature extraction module through a feedback channel, for updating the noise filtering threshold of the dynamic feature vector.

[0012] Further, the bifacial photovoltaic module of the present application comprises:

[0013] a front irradiance sensor and a back reflectance sensor, respectively arranged on the front and back surfaces of the bifacial module, for collecting direct light and reflected light intensity, and transmitting the collected direct light and reflected light intensity data to the dynamic feature extraction module through a data bus;

[0014] a temperature sensor group distributed along the edge and center point of the photovoltaic array for measuring the temperature gradient of the component surface and inputting the temperature gradient data to the dynamic feature extraction module through a bus protocol;

[0015] a monitoring unit connected to the battery cluster of the energy storage system for obtaining the state of charge distribution and charge / discharge rate of each cluster and inputting the state of charge distribution and charge / discharge rate data to the dynamic feature extraction module through a communication interface.

[0016] Further, the photovoltaic assembly of the present application, the prediction model module comprises:

[0017] inputting the direct light and reflected light intensity data collected by the sensor group into the double-channel input layer of the preset long short-term memory network to generate second-level power generation deviation prediction data;

[0018] inputting the temperature gradient data collected by the temperature sensor group into the preset Kalman filter to construct the state transition equation of the component temperature rise and heat dissipation efficiency and output the temperature rise state correction value to the collaborative optimization control module;

[0019] updating the parameters of the preset energy storage efficiency decay model according to the state of charge distribution and charge / discharge rate data collected by the monitoring unit to generate energy storage efficiency decay prediction results.

[0020] Further, the photovoltaic assembly of the present application, the collaborative optimization control module comprises:

[0021] the second-level control loop receives the second-level power generation deviation prediction data generated by the prediction model module and the real-time state of charge data collected by the monitoring unit, and generates an inverter power correction instruction through a preset adaptive droop control algorithm;

[0022] the minute-level control loop optimizes the matching relationship between the energy storage charge / discharge rate and the maximum power point tracking voltage parameter based on the temperature rise state correction value and the energy storage efficiency decay prediction results output by the prediction model module using a preset dynamic programming algorithm;

[0023] the hour-level control loop generates a clustering frequency modulation capacity distribution proportion table through a preset mixed integer linear programming combined with the grid frequency modulation time period constraint and distributes the clustering frequency modulation capacity distribution proportion table to the energy storage system.

[0024] Further, the photovoltaic assembly of the present application, the instruction execution module comprises:

[0025] an inverter interface unit for converting the inverter power correction instruction generated by the second-level control loop of the collaborative optimization control module into a pulse width modulation signal and collecting the actual output power waveform;

[0026] The energy storage control unit adjusts the switching sequence of the battery cluster according to the energy storage charging and discharging power distribution table generated by the minute-level control loop of the cooperative optimization control module, and synchronously monitors the charging and discharging efficiency;

[0027] The feedback channel feeds back the actual output power waveform collected by the inverter interface unit and the charging and discharging efficiency data monitored by the energy storage control unit to the dynamic feature extraction module, so as to update the noise filtering threshold of the dynamic feature vector.

[0028] In a second aspect, the application provides a photovoltaic system applied to the photovoltaic module, comprising:

[0029] The edge computing node is deployed on the side of the photovoltaic array and is connected to the sensor group of the photovoltaic module. The edge computing node stores the optimization instruction into the control instruction database, receives the front surface irradiance, back surface reflectivity and ambient temperature data collected by the sensor group, performs the timestamp alignment and noise filtering processing of the dynamic feature extraction module in the photovoltaic module, and runs the adaptive droop control algorithm of the second-level control loop in the photovoltaic module.

[0030] The cloud scheduling center is connected to the edge computing node through a redundant communication network, receives the hour-level power generation prediction curve and energy storage efficiency decay prediction result generated by the prediction model module in the photovoltaic module, generates a global frequency modulation capacity allocation target based on the preset grid frequency modulation demand and the preset energy storage leasing agreement constraint, and issues the global frequency modulation capacity allocation target to the hour-level control loop of the photovoltaic module.

[0031] The redundant communication network adopts a double CAN bus and a fiber ring network architecture, transmits the second-level inverter power correction instruction, the minute-level energy storage charging and discharging power distribution table and the hour-level cluster frequency modulation capacity allocation proportion table generated by the cooperative optimization control module in the photovoltaic module to the edge computing node and the energy storage system, and uploads the dynamic feature vector and the actual output power feedback data processed by the edge computing node to the cloud scheduling center.

[0032] Further, the photovoltaic system provided by the application, the redundant communication network comprises:

[0033] The main link error rate is collected, and when the main link error rate exceeds the preset threshold of the double CAN bus, the second-level inverter power correction instruction generated by the cooperative optimization control module is switched to the standby fiber link for transmission;

[0034] During the communication interruption, the unsent optimization instruction generated by the cooperative optimization control module is stored locally, and after the communication is restored, the unsent optimization instruction is transmitted to the energy storage system in the timestamp order;

[0035] The real-time monitored communication delay parameter is input into a minute-level control loop of the collaborative optimization control module, and an optimization cycle length of the minute-level control loop is dynamically adjusted.

[0036] Further, the photovoltaic system provided by the application, the cloud scheduling center comprises:

[0037] The hour-level power generation prediction curve generated by the prediction model module in the photovoltaic assembly and the energy storage health state data fed back by the energy storage control unit in the photovoltaic assembly are received.

[0038] According to the preset power grid frequency modulation demand and the preset energy storage leasing agreement constraint, the frequency modulation task allocation weight of each energy storage cluster is calculated.

[0039] The clustering frequency modulation capacity allocation proportion table is sent to the hour-level control loop of the photovoltaic assembly as an input constraint condition of a mixed integer linear programming model of the hour-level control loop.

[0040] Further, the photovoltaic system provided by the application further comprises a fault processing module, which is configured to:

[0041] When the sensor group of the photovoltaic assembly fails, the historical synchronous data stored by the dynamic feature extraction module is called, and a virtual sensor value is generated in combination with the irradiance spatial mapping model of the adjacent photovoltaic array.

[0042] When the interruption duration of the redundant communication network exceeds a set threshold, the energy storage system is switched to an off-grid operation mode, and non-critical loads are cut off according to the frequency modulation task allocation weight calculated by the cloud scheduling center.

[0043] When the power device of the energy storage control unit in the photovoltaic assembly is detected to overcurrent, the charging and discharging rate is reduced, and the switching frequency of the junction temperature estimation model driven by the temperature rise state correction value output by the prediction model module is adjusted.

[0044] Further, the photovoltaic system provided by the application, the fault processing module is further configured to: after the communication is restored, the time difference between the local cached unsent optimization instruction and the latest instruction sent by the cloud scheduling center is compared, and a preset conflict resolution algorithm is used to update the control instruction database of the edge computing node.

[0045] According to the deviation between the real-time state of charge data fed back through the feedback channel and the virtual estimated value generated based on the constant current and constant voltage charging and discharging sequence, the charging and discharging sequence is restarted to calibrate the energy storage capacity.

[0046] The fault event record is input into a case library, a typical fault coping strategy template is generated through cluster analysis, and the adaptive fault tolerance parameters of the collaborative optimization control module are optimized.

[0047] The present application has beneficial effects.

[0048] The present application generates a standardized feature vector through a multi-source sensor data fusion and dynamic feature extraction module, combines a multi-model collaborative prediction mechanism of a long short-term memory network, a Kalman filter and an energy storage efficiency attenuation model, accurately captures the second-level irradiation fluctuation of a bifacial photovoltaic array, the component temperature rise state and the energy storage capacity attenuation trend; based on a hierarchical optimization control architecture, the adaptive droop algorithm is used in the second-level control loop to realize the millisecond-level power grid frequency fluctuation suppression, the rolling horizon optimization algorithm is used in the minute-level control loop to balance the thermodynamic constraints and the life loss, the mixed integer linear programming is used in the hour-level control loop to plan the frequency modulation capacity distribution, forming a multi-time scale collaborative instruction chain; the redundant communication network switches the heterogeneous link through a double CAN bus and a fiber ring network, and the instruction cache transmission mechanism guarantees the real-time and integrity of the control instruction transmission; the fault handling module combines the virtual sensor value generation, the off-grid load removal and the case library self-learning strategy to improve the robustness of the system in the sensor failure, communication interruption and equipment overcurrent scenarios; the edge computing and cloud scheduling collaborative architecture covers the local real-time control and global resource optimization, the feedback channel dynamically calibrates the data preprocessing parameters, and finally solves the multi-dimensional collaborative control problem of the dynamic fluctuation of the bifacial photovoltaic power generation power and the response lag of the rental energy storage system, and significantly improves the energy utilization rate and the power grid frequency modulation accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed in the embodiments. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on the drawings.

[0050] Figure 1 A flowchart of a photovoltaic module system provided by the present application is shown. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical scheme and advantages of the present application clearer, the following will combine the specific embodiments of the present application and the corresponding drawings to clearly and completely describe the technical scheme of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. The technical scheme provided by each embodiment of the present application will be described in detail below in combination with the drawings. In order to better understand the purpose of the present application, the present application will be further described in detail below.

[0052] In a first aspect, the present application provides a photovoltaic module, comprising:

[0053] A sensor group is used to collect the front irradiance, back reflectance and ambient temperature data of the bifacial photovoltaic array, and to synchronously obtain the state of charge and charging / discharging rate of the energy storage system;

[0054] The photovoltaic module provided by the application provides real-time input for the control system through the multi-dimensional data acquisition and transmission mechanism of the sensor group. The components are described in detail as follows:

[0055] The front irradiance sensor is vertically installed on the surface of the front glass layer of the bifacial module using a multi-spectral probe, covering the visible light to near-infrared band, and capturing the direct light intensity and spectral distribution in real time through a photodiode array. The back reflectance sensor integrates a wide-angle lens and a polarization filter, and is arranged obliquely at the edge of the module back plate to collect the multi-angle incident intensity of the water surface reflected light, and outputs the dynamic reflection coefficient combined with the reflectance calculation model. The two sensors synchronously transmit the light intensity raw data to the input buffer area of the dynamic feature extraction module through the RS-485 bus, solving the time synchronization problem of the front and back irradiation data, and supporting the real-time calculation of the irradiance difference coefficient.

[0056] The temperature sensor group selects a platinum resistance PT1000 probe, which is distributed in a grid shape along the diagonal and center point of the photovoltaic array, and four temperature measurement nodes are arranged per square meter. The temperature gradient data is transmitted to the temperature processing unit of the dynamic feature extraction module through the SPI interface after analog-to-digital conversion, and the transverse and longitudinal temperature difference coefficients of the module surface are calculated to identify the local overheating area caused by the fish-light complementary environment and provide the thermodynamic state input for the Kalman filter.

[0057] The monitoring unit includes a high-precision coulomb meter and a Hall current sensor, which are embeddedly installed in the bus cabinet of the energy storage system battery cluster. The coulomb meter measures the state of charge of each cluster based on the charge integration method, and the Hall sensor non-contactly collects the charging and discharging current waveform through the magnetic balance principle. The monitoring data is transmitted to the energy storage data processing unit of the dynamic feature extraction module through the CAN bus to calculate the state of charge dispersion and charging / discharging rate fluctuation between the battery clusters in real time, and provide dynamic input for the parameter update of the energy storage efficiency decay model.

[0058] The data synchronization and transmission of the sensors are realized through heterogeneous communication protocols. The front and back light sensor data are aligned by millisecond timestamp through the RS-485 bus to eliminate the time deviation caused by the difference in sampling clock; the SPI interface of the temperature sensor group ensures the low-delay transmission of high-precision temperature gradient data; and the CAN bus protocol of the monitoring unit supports the parallel collection of multi-node energy storage state data. After the normalization of the multi-source data, the dynamic feature extraction module generates a dynamic feature vector including the irradiance difference, temperature fluctuation and energy storage state parameters to drive the collaborative optimization of the subsequent prediction model and control algorithm.

[0059] In the technical solution of the sensor group, the spectral and polarization characteristics of the front and back light sensors are designed to solve the data acquisition accuracy problem of the double-sided assembly in a complex reflection environment; the grid layout of the temperature sensor and the four-wire measurement method improve the spatial resolution of the surface heat distribution of the assembly; the non-contact acquisition mechanism of the energy storage monitoring unit avoids interference with the operation state of the battery cluster. The multi-bus protocol data transmission architecture realizes the time sequence consistency of heterogeneous data sources through timestamp alignment and buffer management, provides standardized multi-dimensional input for the dynamic feature extraction module, and supports the control system from second-level response to hour-level scheduling for full-cycle optimization requirements.

[0060] The dynamic feature extraction module is connected to the sensor group, and the collected data is timestamped and noise filtered, generating a dynamic feature vector including irradiance difference, temperature fluctuation and energy storage state parameters, and outputting the dynamic feature vector to the prediction model module.

[0061] In the photovoltaic module described in the application, the dynamic feature extraction module converts multi-source sensor data into a feature vector that can be analyzed by the control model through a standardized data processing flow. The following describes each technical step in detail:

[0062] The timestamp alignment unit synchronizes the sampling clock of the sensor group based on the network time protocol, eliminating the time offset caused by the difference in acquisition frequency of heterogeneous sensors. When the direct light and reflected light intensity data are transmitted through the RS-485 bus, a millisecond timestamp is added; when the temperature gradient data is transmitted through the SPI interface, the acquisition time is aligned using a hardware interrupt mechanism; and the energy storage state data is synchronized through the CAN bus protocol timestamp field. The aligned data stream is stored in a ring buffer according to the preset time window length, forming a multi-dimensional data set with consistent timing.

[0063] The noise filtering module uses a wavelet transform and moving average combination algorithm to perform hierarchical denoising on the original data. For the instantaneous mutation characteristics of irradiance data (such as cloud cover), the high-frequency components are retained to capture second-level fluctuations; for baseline drift in temperature gradient data (such as ambient temperature variation), low-frequency noise is filtered out through wavelet decomposition; and for the charge and discharge rate pulse interference in the energy storage state parameter, a sliding window mean filter is used for smoothing. The processed data is stored in a time series database according to the sensor type, outputting standardized parameters such as irradiance difference coefficient, temperature fluctuation variance and state of charge dispersion.

[0064] The dynamic feature vector generation unit fuses the preprocessed multi-source data according to preset weight coefficients. The irradiance difference coefficient is calculated by the front-back light intensity ratio, reflecting the irradiation asymmetry in the double-sided power generation scene; the temperature fluctuation variance is calculated based on the standard deviation of the gridized temperature measurement nodes, representing the uniformity of the component thermal distribution; the state of charge dispersion is quantified by the maximum difference of the SOC between the battery clusters, indicating the consistency level of the energy storage system. The fused feature vector is pushed to the input interface of the prediction model module through the message queue, driving the model inference of the long short-term memory network and Kalman filter.

[0065] The logical relationship between the modules is as follows: timestamp alignment provides time-consistent data input for noise filtering, eliminating feature calculation errors caused by collection time deviation; the layered denoising algorithm retains valid information for different data types, improving the signal-to-noise ratio of the feature vector; the multi-parameter fusion mechanism maps heterogeneous data into a unified dimension of the control model input, supporting the generation of subsequent prediction and optimization instructions. The output of the dynamic feature extraction module is interconnected with the prediction model module and the collaborative optimization control module through the data bus, forming a complete link from data preprocessing to control decision, meeting the requirements of real-time and accuracy of industrial process control systems.

[0066] The prediction model module receives the front irradiance, back reflectivity and temperature fluctuation data in the dynamic feature vector, models the second-level irradiance change of the front irradiance using a preset long short-term memory network, generates second-level power generation power deviation prediction data, simultaneously inputs the temperature fluctuation data into a preset Kalman filter for component temperature rise state evaluation, outputs a temperature rise state correction value, and updates the preset energy storage efficiency decay model parameters according to the state of charge and charge-discharge rate data, generating energy storage efficiency decay prediction results;

[0067] The prediction model module generates dynamic prediction results through a multi-model collaborative processing mechanism, providing input references for layered control. The following describes the technical steps in detail:

[0068] The double-channel input layer of the long short-term memory network receives the front irradiance and back reflected light intensity time series data in the dynamic feature vector. The direct light channel uses a sliding window to extract second-level irradiance mutation features, with a window length of 5 seconds, covering the instantaneous fluctuations caused by rapid cloud cover; the reflected light channel calculates the dynamic reflectivity contribution value through a time decay weight algorithm, eliminating random noise interference from water surface ripple reflection. After fusion by the gated recurrent unit, a power generation power deviation prediction sequence for the next 5 seconds is generated, the prediction results are matched with the real-time power sampling values through the timestamp alignment module, forming a deviation correction reference for the second-level control loop, supporting the rapid adjustment of the inverter power output.

[0069] The Kalman filter constructs a state transition equation based on a component heat conduction model and inputs temperature gradient data in a dynamic characteristic vector. The state variables include the internal temperature rise rate of the component and the heat dissipation efficiency coefficient, and the observation variables are the surface temperature gradients collected by the temperature sensor group. The process noise matrix dynamically adjusts the noise covariance according to the environmental temperature fluctuation, and the observation noise matrix is calibrated by the measurement accuracy of the temperature sensor. Through the iterative prediction and update steps, the temperature rise state correction value of the internal hot spot area of the component is output, the heat dissipation strategy optimization instruction is generated by combining the heat sink control parameters, and the matching relationship of the charging and discharging rate and the heat dissipation fan speed is adjusted in the minute-level control loop.

[0070] The energy storage efficiency attenuation model calculates the battery cluster capacity attenuation gradient and the historical correlation of the charging and discharging rate according to the state of charge distribution data in the dynamic characteristic vector. The model parameters are updated online through an incremental learning mechanism, and during the calibration process, a sliding window is introduced to filter effective data samples and eliminate distorted data caused by abnormal charging and discharging events (such as overvoltage protection triggering). The updated model outputs the energy storage efficiency attenuation prediction curve, and the prediction result is packaged as a charging and discharging power allocation constraint condition for the minute-level control loop, limiting the charging and discharging depth of high-loss battery clusters and balancing energy utilization and device life loss.

[0071] The data interaction between the models is realized through a standardized interface. The generated power deviation prediction data is input into the second control loop after normalization, driving the adaptive droop control algorithm; the temperature rise correction value and the efficiency attenuation prediction result are input into the dynamic programming algorithm as a set of constraint parameters, optimizing the minute-level charging and discharging strategy; the energy storage health state data and the frequency modulation demand are fused in the cloud scheduling center to generate a hourly-level cluster frequency modulation capacity allocation proportion table. The output parameters of the prediction model module are transmitted through a data bus in layers, forming a full-cycle control link from the second-level response to the hourly-level scheduling, meeting the multi-time scale coordination needs of industrial process control.

[0072] The cooperative optimization control module is connected to the prediction model module and includes a second control loop, a minute-level control loop, and a hourly-level control loop. The second control loop is input with the second control loop power deviation prediction data, and an inverter power correction instruction is generated through a preset adaptive droop control algorithm. The temperature rise state correction value and the energy storage efficiency attenuation prediction result are input into the minute-level control loop, a charging and discharging power allocation table is generated by combining a preset rolling horizon optimization algorithm, and a grid frequency modulation period constraint is input into the hourly-level control loop, and a cluster frequency modulation capacity allocation proportion table is generated through a preset mixed integer linear programming;

[0073] The cooperative optimization control module in the photovoltaic module realizes dynamic matching and optimization of multi-time scale instructions through a hierarchical control architecture. The following describes each control loop in detail:

[0074] The second-level control ring receives the second-level power generation deviation prediction data output by the prediction model module and the real-time state of charge data collected by the monitoring unit, and generates an inverter power correction instruction through an adaptive droop control algorithm. In the algorithm, a dynamic mapping relationship between the droop coefficient and the state of charge is established. When a sudden change in irradiance is detected, the inverter output limit is adjusted in real time according to the current state of charge of the energy storage system. The power compensation is preferentially performed by the battery cluster with low state of charge dispersion. The correction instruction is converted into a pulse width modulation signal and then sent to the inverter drive circuit to achieve millisecond-level response and suppress transient fluctuations in grid frequency.

[0075] The minute-level control ring uses a rolling horizon optimization algorithm to construct a multi-stage decision model based on the temperature rise state correction value output by the Kalman filter and the energy storage efficiency decay prediction curve. The algorithm divides the time window into discrete state nodes. At each node, the optimal matching relationship between the charging and discharging rate and the maximum power point tracking voltage is calculated. The temperature rise correction value is used as a thermodynamic constraint, and the efficiency decay prediction value is used as a life loss constraint to solve the charging and discharging power distribution path that meets the multi-objective optimization. The calculation result is packaged as a power distribution table and transmitted to the energy storage converter through the CAN bus to dynamically adjust the battery cluster switching order and equalization strategy.

[0076] The hour-level control ring integrates the grid frequency regulation demand curve and the energy storage leasing agreement time period constraint. In the mixed integer linear programming model, the frequency regulation capacity, battery life loss coefficient, and dispatching time period are defined as decision variables. The objective function maximizes the frequency regulation income, and the constraint conditions include the remaining capacity of the energy storage system, the upper limit of the charging and discharging cycle number, and the service time period agreed in the agreement. After the model is solved, a clustered frequency regulation capacity distribution proportion table is generated. The frequency regulation task weight of each battery cluster is divided by time granularity, and the table is sent to the energy storage system main controller through the fiber communication network to guide the frequency regulation task execution in the future hours.

[0077] The control rings achieve instruction coordination and parameter feedback through the data bus. The real-time response result of the second-level control ring provides dynamic state input for the minute-level optimization. The charging and discharging power distribution table output by the minute-level optimization serves as the boundary condition for the hour-level scheduling model, limiting the solution space of the long-term frequency regulation capacity distribution scheme. The clustered frequency regulation capacity distribution proportion table calibrates the state of charge mapping parameters of the second-level control ring through the feedback channel, forming a closed-loop optimization link from short-term response to long-term planning. The redundant communication network synchronously transmits the second-level inverter instruction, the minute-level power distribution table, and the hour-level frequency regulation proportion table. The edge computing node performs local real-time control, and the cloud scheduling center coordinates global resource allocation, achieving dynamic balance between control accuracy and system economy.

[0078] The instruction execution module receives the inverter power correction instruction, the energy storage charge and discharge power distribution table and the cluster frequency modulation capacity distribution proportion table, adjusts the maximum power point tracking voltage parameter of the double-sided photovoltaic array and the charge and discharge rate of the energy storage system, and transmits the adjusted parameter of the double-sided photovoltaic array and the actual output power data to the dynamic feature extraction module through a feedback channel, for updating the noise filtering threshold of the dynamic feature vector.

[0079] The photovoltaic module described in the application realizes closed-loop optimization through control instruction conversion and feedback mechanism. The following will be described in detail:

[0080] The inverter interface unit receives the inverter power correction instruction generated by the cooperative optimization control module, and converts the digital instruction into a duty cycle adjustable modulation signal through a pulse width modulation controller. The modulation signal is amplified by an isolation drive circuit and transmitted to the inverter power switching device, to adjust the DC bus voltage and AC output waveform. The synchronous acquisition unit integrates a Hall sensor and a high-speed analog-to-digital conversion module, to capture the voltage and current waveform of the inverter output in real time, calculate the actual output power instantaneous value, and upload to the data processing unit through a serial peripheral interface, to verify the execution effect of the control instruction.

[0081] The energy storage control unit analyzes the energy storage charge and discharge power distribution table issued under the minute-level control ring, extracts the charge and discharge rate, voltage threshold and time window parameter. The multi-channel relay array dynamically adjusts the switching order according to the state of charge dispersion and health state score of the battery cluster, and preferentially calls the battery cluster with high consistency to perform the task. The charge and discharge efficiency monitoring module calculates the energy conversion loss rate in real time through the coulomb counting method, corrects the efficiency calculation model combined with the battery surface temperature data, generates an efficiency monitoring report and stores it in the local cache area, to provide an input reference for subsequent optimization iteration.

[0082] The feedback channel configures a data compression and verification module, encapsulates the inverter actual output power waveform data and the energy storage charge and discharge efficiency monitoring report into a standardized data packet. The data packet is returned to the input interface of the dynamic feature extraction module through an industrial Ethernet, and the noise filtering unit dynamically adjusts the threshold parameter of wavelet denoising based on the waveform statistical characteristics (such as harmonic distortion rate and peak fluctuation amount). The updated filtering parameter is applied to the preprocessing of sensor data in the next period, to form a closed-loop optimization link from instruction execution to parameter calibration.

[0083] The logical association between the modules is as follows: the signal conversion and collection mechanism of the inverter interface unit ensures the accurate execution of the control instructions, and the actual waveform monitoring verifies the response effect; the switching strategy and efficiency monitoring of the energy storage control unit realize the dynamic optimization of energy scheduling; the data back transmission and threshold updating mechanism of the feedback channel form a closed-loop calibration. The adjusted maximum power point tracking voltage parameter drives the photovoltaic array output voltage adjustment through the pulse width modulation signal, and the charge and discharge rate adjustment instruction controls the battery cluster switching through the relay array, both of which cooperate to maintain the stability of the system power output. The actual output data is transmitted back to the dynamic feature extraction module through the feedback channel, driving the iterative update of the noise filtering threshold, improving the accuracy of subsequent data preprocessing, and supporting the continuous optimization of system adaptability.

[0084] The application relates to a photovoltaic module based on a multi-sensor data fusion and hierarchical optimization control architecture, which realizes dynamic cooperative regulation of a double-sided photovoltaic array and an energy storage system. The steps of the technical scheme are described in detail as follows:

[0085] The sensor group is deployed on the front and back surfaces of the double-sided photovoltaic array and the energy storage system side. The front irradiance sensor adopts a multispectral probe vertically installed, covering the visible light to near-infrared band, and capturing the direct light intensity and spectral distribution in real time. The back reflectivity sensor integrates a wide-angle lens and a polarization filter, and is obliquely arranged at the edge of the module back plate to measure the reflected light intensity at multiple angles. The temperature sensor group selects a platinum resistance PT1000 probe, which is distributed in a grid shape along the diagonal and center point of the photovoltaic array, and eliminates the wire resistance error through four-wire measurement to collect the temperature gradient data of the module surface. The energy storage system side is configured with a high-precision coulomb meter and a Hall current sensor to non-contact monitor the state of charge distribution and charge and discharge rate of the battery cluster. Each sensor transmits the original data to the dynamic feature extraction module through industrial Ethernet and CAN bus to form a multi-source heterogeneous data collection layer.

[0086] The dynamic feature extraction module integrates a time series database and a signal processing unit to perform standardized preprocessing on the sensor data. The timestamp alignment unit synchronizes the sampling clocks of each sensor based on the network time protocol, eliminating the time offset of heterogeneous data; the noise filtering process adopts a wavelet transform and moving average combination algorithm to retain the high-frequency mutation characteristics of irradiance data and correct the baseline drift of temperature data. The processed dynamic feature vector includes the irradiance difference coefficient, temperature fluctuation variance and energy storage state parameters, which are pushed to the prediction model module through the message queue to complete the data feature engineering.

[0087] In the prediction model module, the dual-channel input layer of the long short-term memory network receives the front and back irradiance time series data, captures the second-level irradiance mutation characteristics through the gated recurrent unit, and outputs the power deviation prediction sequence in the next 5 seconds. The Kalman filter constructs a component heat conduction state space model based on the temperature gradient data, iteratively corrects the heat dissipation efficiency parameter, and outputs the temperature rise state correction value. The energy storage efficiency decay model fits the capacity decay curve according to the state of charge history data, dynamically updates the model parameters and generates the efficiency decay prediction results. The prediction data is input into different levels of control loops through the data bus to support multi-time scale decision-making.

[0088] The cooperative optimization control module is divided into second-level, minute-level and hour-level control loops. The second-level control loop adopts an adaptive droop control algorithm, dynamically adjusts the inverter power output limit value according to the power deviation prediction value and real-time state of charge, preferentially calls the high-response-rate energy storage unit for power compensation, and suppresses the grid frequency fluctuation. The minute-level control loop is based on a rolling horizon optimization algorithm, takes the temperature rise correction value and efficiency decay prediction as constraint conditions, solves the optimal matching sequence of energy storage charging and discharging rate and maximum power point tracking voltage through dynamic programming, and generates a charging and discharging power distribution table. The hour-level control loop integrates the grid frequency modulation period demand and the energy storage leasing agreement, adopts a mixed integer linear programming model to calculate the cluster frequency modulation capacity distribution proportion table, and generates a frequency modulation instruction sequence according to the time granularity.

[0089] In the instruction execution module, the inverter interface unit converts the power correction instruction into a pulse width modulation signal to drive the IGBT switching device to adjust the output voltage waveform; the energy storage control unit analyzes the charging and discharging power distribution table, controls the parallel switching sequence of the battery cluster, and synchronously monitors the charging and discharging efficiency. The feedback channel collects the actual output power waveform and efficiency data, calculates the noise filter threshold update value through the sliding window statistical method, and returns to the noise filter parameter configuration unit of the dynamic feature extraction module to form a closed-loop parameter optimization mechanism.

[0090] The edge computing node is deployed at the photovoltaic array side to perform local data preprocessing and second-level control algorithm, reducing the cloud communication delay; the cloud scheduling center generates a global frequency modulation capacity distribution target through mixed integer linear programming based on the hour-level power generation prediction curve and the energy storage health status data. The redundant communication network adopts a heterogeneous architecture of dual CAN bus and fiber ring network, the main link transmits the second-level instruction, the standby link seamlessly switches when the bit error rate exceeds the standard, and the reliability of instruction transmission is ensured.

[0091] The fault processing module calls historical synchronous data and adjacent array irradiance mapping model to generate virtual sensor value when the sensor fails; switches to off-grid operation mode after communication interruption exceeds threshold, and removes non-critical load according to frequency task weight; dynamically adjusts switching frequency combined with temperature rise correction value when power device overflows.

[0092] Each technical module is interconnected with a communication network through a data bus, forming a closed-loop control link of "perception-prediction-optimization-execution-feedback". Sensor data is processed by dynamic feature extraction and prediction model, driving the generation of optimized instructions by hierarchical control loop; the execution results of the instructions are calibrated by feedback channel to realize the adaptability of the system; the collaborative architecture of edge computing and cloud scheduling covers multi-time scale control requirements, and the redundant communication and fault processing mechanism enhances the robustness of the system. The scheme effectively solves the cooperative control problem of dynamic fluctuation of double-sided photovoltaic power generation and response lag of energy storage, and improves the energy utilization rate and grid frequency modulation accuracy.

[0093] Specifically, the photovoltaic module described in the application comprises:

[0094] The front irradiance sensor and the back reflectivity sensor are respectively arranged on the front and back surfaces of the double-sided module to collect direct light and reflected light intensity, and transmit the collected direct light and reflected light intensity data to the dynamic feature extraction module through the data bus.

[0095] The temperature sensor group is distributed along the edge and center point of the photovoltaic array to measure the temperature gradient of the module surface, and input the temperature gradient data to the dynamic feature extraction module through the bus protocol.

[0096] The monitoring unit is connected to the battery cluster of the energy storage system to obtain the state of charge distribution and charge / discharge rate of each cluster, and input the state of charge distribution and charge / discharge rate data to the dynamic feature extraction module through the communication interface.

[0097] In the photovoltaic module described in the application, the sensor group provides standardized input for the dynamic feature extraction module through multi-dimensional data acquisition and transmission mechanism. The components are described in detail as follows:

[0098] The front irradiance sensor adopts a multispectral probe vertically installed on the surface of the front glass layer of the double-sided assembly, covering the visible light to near-infrared band, and captures the direct light intensity and spectral distribution in real time through a photodiode array. The back reflectivity sensor integrates a wide-angle lens and a polarizing filter, and is obliquely arranged at the edge of the assembly back plate to collect the multi-angle incident intensity of the water surface reflected light, and outputs the dynamic reflection coefficient combined with the reflectivity calculation model. The two groups of sensors transmit the light intensity raw data synchronously to the input buffer area of the dynamic feature extraction module through the RS-485 bus, solving the time synchronization problem of the front and back irradiance data, and supporting the real-time calculation of the irradiance difference coefficient.

[0099] The temperature sensor group selects a platinum resistance PT1000 probe, which is distributed in a grid shape along the diagonal and center point of the photovoltaic array, and four temperature measurement nodes are arranged per square meter. The temperature gradient data is transmitted to the temperature processing unit of the dynamic feature extraction module through the SPI interface after analog-to-digital conversion, and the transverse and longitudinal temperature difference coefficients of the assembly surface are calculated to identify the local overheating area caused by the fish-light complementary environment and provide the thermodynamic state input for the Kalman filter.

[0100] The monitoring unit includes a high-precision coulomb meter and a Hall current sensor, which are embeddedly installed in the busbar cabinet of the energy storage system battery cluster. The coulomb meter measures the state of charge of each cluster based on the charge integration method, and the Hall sensor non-contactly collects the charge and discharge current waveform through the magnetic balance principle. The monitoring data is transmitted to the energy storage data processing unit of the dynamic feature extraction module through the CAN bus, and the state of charge dispersion and charge and discharge rate fluctuation between the battery clusters are calculated in real time to provide dynamic input for the parameter update of the energy storage efficiency decay model.

[0101] Each group of sensors realizes data synchronization and transmission through heterogeneous communication protocols. The front and back light sensor data are aligned by millisecond-level time stamp through the RS-485 bus, eliminating the timing deviation caused by the difference in sampling clock; the SPI interface of the temperature sensor group ensures low-delay transmission of high-precision temperature gradient data; and the CAN bus protocol of the monitoring unit supports parallel collection of multi-node energy storage state data. After the dynamic feature extraction module normalizes the multi-source data, it generates a dynamic feature vector including irradiance difference, temperature fluctuation and energy storage state parameters, which drives the collaborative optimization of the subsequent prediction model and control algorithm.

[0102] The multi-dimensional layout of the sensor group and the standardized data transmission mechanism provide basic data support for the light, heat and electrical state monitoring of the photovoltaic module. The time synchronization of the front and back surface irradiation data and the dynamic calculation of the reflectivity solve the problem of irradiance difference evaluation in the double-sided power generation scene; the grid distribution and high-precision measurement of the temperature sensor identify the spatial non-uniformity of the component temperature rise; the non-contact acquisition method of the energy storage monitoring unit avoids interference with the operation state of the battery cluster. The fusion processing of the three groups of data by the dynamic feature extraction module constructs the perception layer input of the closed-loop control system, and meets the requirements of data integrity and real-time of industrial process control.

[0103] Specifically, the photovoltaic module provided by the application comprises a prediction model module.

[0104] The direct light and reflected light intensity data collected by the sensor group are input into the double-channel input layer of the preset long short-term memory network to generate second-level power generation power deviation prediction data;

[0105] The temperature gradient data collected by the temperature sensor group are input into the preset Kalman filter to construct the state transition equation of the component temperature rise and heat dissipation efficiency, and output the temperature rise state correction value to the collaborative optimization control module;

[0106] According to the state of charge distribution and charge and discharge rate data collected by the monitoring unit, the parameters of the preset energy storage efficiency attenuation model are updated to generate an energy storage efficiency attenuation prediction result.

[0107] In the photovoltaic module provided by the application, the prediction model module generates dynamic prediction results through a multi-model collaborative processing mechanism to provide input reference for hierarchical control. The following describes the technical steps in detail:

[0108] The double-channel input layer of the long short-term memory network receives the front surface irradiance and back surface reflected light intensity time series data respectively. The direct light channel adopts a sliding window to extract second-level irradiation mutation features, and the reflected light channel calculates dynamic reflectivity contribution values through time decay weights. After fusion by the gated recurrent unit, the double-channel output generates a power generation power deviation prediction sequence within the next 5 seconds, capturing the instantaneous power fluctuations caused by cloud cover or water surface reflection fluctuations. The prediction data are matched with real-time power sampling values through a time stamp alignment module to form a second-level control loop deviation correction reference, supporting the rapid adjustment of inverter power output.

[0109] The Kalman filter constructs a state transition equation based on a component heat conduction model and inputs temperature sensor group collected grid temperature gradient data. By iteratively calculating the coupling relationship between the temperature rise rate and the heat dissipation efficiency, the temperature distribution parameters of the hot spot area in the component are dynamically corrected. The process noise matrix is dynamically adjusted according to the environmental temperature fluctuation, and the observation noise matrix is calibrated by the measurement accuracy of the temperature sensor. The output results include the temperature rise state correction value and the heat dissipation efficiency evaluation parameter, which are input into the minute-level control loop to optimize the dynamic matching of the charging and discharging rate and the heat dissipation strategy.

[0110] The energy storage efficiency attenuation model calculates the capacity attenuation gradient and the historical correlation of the charging and discharging rate according to the state of charge distribution data collected by the monitoring unit. The model parameters are updated online through an incremental learning mechanism, and effective data samples are selected by introducing a sliding window in the calibration process to eliminate the interference of abnormal charging and discharging events on the attenuation prediction. The updated model outputs the energy storage efficiency attenuation prediction curve as a charging and discharging power allocation constraint condition for the minute-level control loop, balancing energy utilization and device life loss.

[0111] In the prediction model module, the long short-term memory network focuses on the influence of second-level light fluctuation on power generation, the Kalman filter solves the spatial non-uniformity evaluation of the temperature rise state, and the energy storage efficiency attenuation model quantifies the performance degradation trend of the battery. Three of them establish prediction models for light, heat and electrical parameters respectively, and the output data is processed by the hierarchical processing of the collaborative optimization control module to form a multi-level control instruction with second-level fast response, minute-level optimization matching and hour-level overall scheduling. The prediction results and real-time monitoring data are time-stamped in the time stamp alignment module to ensure the dynamic accuracy of the control instruction.

[0112] The data interaction between the models is realized through a standardized interface. The direct light and reflected light prediction data are normalized and input into the second-level control loop; the temperature rise correction value and the efficiency attenuation prediction result are packaged as a constraint parameter set to drive the minute-level optimization algorithm; the energy storage health state data and the frequency modulation demand are fused in the cloud scheduling center to generate a hour-level capacity allocation scheme. The output parameters of each prediction model are transmitted through a data bus in layers to form a whole cycle control link from instantaneous response to long-term planning, meeting the multi-time scale collaborative needs of industrial process control.

[0113] Specifically, the photovoltaic module described in the application comprises a collaborative optimization control module, which comprises:

[0114] The second-level control loop receives the second-level power generation deviation prediction data generated by the prediction model module and the real-time state of charge data collected by the monitoring unit, and generates an inverter power correction instruction through a preset adaptive droop control algorithm;

[0115] The minute-level control ring adopts a preset dynamic programming algorithm to optimize the matching relationship between the energy storage charge and discharge rate and the maximum power point tracking voltage parameter based on the temperature rise state correction value and the energy storage efficiency attenuation prediction result output by the prediction model module.

[0116] The hour-level control ring generates a clustering frequency modulation capacity allocation proportion table through a preset mixed integer linear programming combined with grid frequency modulation period constraints and issues the clustering frequency modulation capacity allocation proportion table to the energy storage system.

[0117] The photovoltaic module disclosed in the application realizes multi-time scale instruction generation and dynamic matching through a hierarchical control architecture by the synergistic optimization control module.

[0118] The second-level control ring receives the power generation power deviation prediction data output by the long short-term memory network and the real-time state of charge data collected by the monitoring unit, and generates an inverter power correction instruction through an adaptive droop control algorithm. In the algorithm, a dynamic mapping relationship between the droop coefficient and the state of charge is established. When a power deviation caused by a sudden change in irradiance is detected, the inverter output limit value is adjusted in real time according to the current state of charge of the energy storage system, and the battery cluster with low state of charge dispersion is preferentially called to perform power compensation. The correction instruction is converted into a pulse width modulation signal and then issued to the inverter drive circuit to achieve a millisecond-level response and suppress transient fluctuations in the grid frequency.

[0119] The minute-level control ring adopts a dynamic programming algorithm to construct a multi-stage optimization model based on the temperature rise state correction value output by the Kalman filter and the energy storage efficiency attenuation prediction curve. The algorithm divides the time window into discrete state nodes, calculates the matching relationship between the charge and discharge rate and the maximum power point tracking voltage at each node, and solves the charge and discharge power distribution path that meets the multi-objective optimization by taking the temperature rise correction value as the thermodynamic constraint and the efficiency attenuation prediction value as the life loss constraint. The calculation result is packaged as a power distribution table and transmitted to the energy storage converter through the CAN bus to dynamically adjust the battery cluster switching sequence and the equalization strategy.

[0120] The hour-level control ring integrates the grid frequency modulation demand curve and the energy storage leasing agreement period constraint, and defines the frequency modulation capacity, battery life loss coefficient and scheduling period as decision variables in the mixed integer linear programming model. The objective function takes the maximum frequency modulation income as the core, and the constraint conditions include the remaining capacity of the energy storage system, the upper limit of the charge and discharge cycle number and the service period agreed in the agreement. After the model is solved, a clustering frequency modulation capacity allocation proportion table is generated, the frequency modulation task weight of each battery cluster is divided according to the time granularity, and the table is issued to the energy storage system main controller through the optical fiber communication network to guide the frequency modulation task execution in the future several hours.

[0121] Each control loop realizes instruction cooperation through a data bus. The fast response of the second-level control loop provides real-time state input for the minute-level optimization, the minute-level optimization result is used as a boundary condition to restrict the solution space of the hour-level scheduling model, and the hour-level distribution table is used to calibrate the state of charge mapping parameter of the second-level control loop through a feedback channel. The hierarchical control architecture covers the whole cycle demand from the millisecond-level transient response to the hour-level resource scheduling, and forms a dynamic closed-loop optimization link.

[0122] The dynamic adjustment mechanism of the droop coefficient of the second-level control loop is complementary to the charge-discharge path optimization of the minute-level control loop, the former guarantees the instantaneous stability of the power grid frequency, and the latter balances the device life and energy efficiency; the capacity distribution table of the hour-level control loop provides a long-term constraint framework for the minute-level optimization, and avoids global resource imbalance caused by local optimization. The three-level control instructions are transmitted hierarchically through a redundant communication network, the edge computing node executes the second-level instructions, the cloud scheduling center coordinates the hour-level strategy, and the control precision and system reliability are improved coordinately.

[0123] Specifically, the photovoltaic module provided by the application comprises:

[0124] The inverter interface unit converts the inverter power correction instruction generated by the second-level control loop of the cooperative optimization control module into a pulse width modulation signal, and collects an actual output power waveform;

[0125] The energy storage control unit adjusts the switching order of the battery cluster according to the energy storage charge-discharge power distribution table generated by the minute-level control loop of the cooperative optimization control module, and synchronously monitors the charge-discharge efficiency;

[0126] The feedback channel returns the actual output power waveform collected by the inverter interface unit and the charge-discharge efficiency data monitored by the energy storage control unit to the dynamic feature extraction module, so as to update the noise filtering threshold of the dynamic feature vector.

[0127] In the photovoltaic module provided by the application, the instruction execution module realizes closed-loop optimization through a control instruction conversion and feedback mechanism. The following will be described in detail:

[0128] The inverter interface unit receives the inverter power correction instruction generated by the second-level control loop, and converts the digital instruction into a duty cycle adjustable modulation signal through a pulse width modulation controller. The modulation signal is amplified through an isolation driving circuit and then transmitted to an inverter power switching device, so as to adjust the direct current bus voltage and alternating current output waveform. The synchronous acquisition unit integrates a Hall sensor and a high-speed analog-to-digital conversion module, and can capture the voltage and current waveforms of the inverter output in real time, calculate the actual output power instantaneous value, and upload the actual output power instantaneous value to the data processing unit through a serial peripheral interface, so as to verify the execution effect of the control instruction.

[0129] The energy storage control unit analyzes the energy storage charging and discharging power distribution table issued under the minute-level control ring, extracts the charging and discharging rate, voltage threshold and time window parameters. The multi-channel relay array dynamically adjusts the switching order according to the state of charge dispersion and health state score of the battery cluster, and preferentially calls the battery cluster with high consistency to perform the task. The charging and discharging efficiency monitoring module calculates the energy conversion loss rate in real time through the coulomb counting method, corrects the efficiency calculation model combined with the battery surface temperature data, generates an efficiency monitoring report and stores it in the local cache area, providing an input benchmark for subsequent optimization iteration.

[0130] The feedback channel configuration data compression and verification module encapsulates the actual output power waveform data of the inverter and the energy storage charging and discharging efficiency monitoring report into a standardized data packet. The data packet is returned to the input interface of the dynamic feature extraction module through the industrial Ethernet, and the noise filtering unit dynamically adjusts the threshold parameters of wavelet denoising based on waveform statistical features such as harmonic distortion rate and peak fluctuation. The updated filtering parameters are applied to the preprocessing of sensor data in the next cycle, forming a closed-loop optimization link from instruction execution to parameter calibration.

[0131] The signal conversion and acquisition mechanism of the inverter interface unit ensures the accurate execution of control instructions, and verifies the response effect through actual waveform monitoring; the switching strategy and efficiency monitoring of the energy storage control unit realize the dynamic optimization of energy scheduling; the data return and threshold updating mechanism of the feedback channel form a closed-loop calibration. The synergistic effect of the three converts the layered control instructions into device actions, and continuously optimizes the data processing accuracy through real-time feedback, supporting the improvement of system adaptability.

[0132] The data flow and logical relationship between the components are as follows: the second-level control instructions drive device actions through the inverter interface unit, and the actual output data is returned through the feedback channel; the minute-level charging and discharging strategy drives the energy storage control unit to adjust the operating state, and the efficiency data is fed back to the cloud scheduling center simultaneously; the dynamic feature extraction module updates the noise filtering parameters according to the feedback data, improves the data preprocessing accuracy of the subsequent data, and forms a closed-loop control link of "instruction generation - execution verification - parameter optimization". Through multi-level data interaction and parameter iteration, the architecture realizes efficient cooperative control of photovoltaic components and energy storage systems.

[0133] In the second aspect, referring to Figure 1 The application provides a photovoltaic system applied to the photovoltaic component, which comprises:

[0134] An edge computing node is deployed at the side of the photovoltaic array, connected with a sensor group of the photovoltaic module, the edge computing node stores optimization instructions into a control instruction database, receives front side irradiance, back side reflectance and ambient temperature data collected by the sensor group, performs timestamp alignment and noise filtering processing of a dynamic feature extraction module in the photovoltaic module, and runs an adaptive droop control algorithm of a second-level control loop of the photovoltaic module.

[0135] A cloud scheduling center is connected with the edge computing node through a redundant communication network, receives an hour-level power generation prediction curve and energy storage efficiency decay prediction result generated by a prediction model module in the photovoltaic module, generates a global frequency modulation capacity allocation target based on a preset power grid frequency modulation demand and a preset energy storage leasing agreement constraint, and issues the global frequency modulation capacity allocation target to the hour-level control loop of the photovoltaic module.

[0136] A redundant communication network adopts a double CAN bus and a fiber ring network architecture, transmits second-level inverter power correction instructions, minute-level energy storage charging and discharging power allocation tables and hour-level cluster frequency modulation capacity allocation proportion tables generated by a cooperative optimization control module in the photovoltaic module to the edge computing node and the energy storage system, and uploads dynamic feature vectors and actual output power feedback data processed by the edge computing node to the cloud scheduling center.

[0137] The photovoltaic system disclosed in the application realizes multi-level control optimization through edge computing and cloud cooperative architecture. The following describes the technical components in detail:

[0138] The edge computing node is deployed at the side of the photovoltaic array, and collects front side irradiance, back side reflectance and ambient temperature data through Modbus protocol connection with a sensor group. A timestamp alignment unit synchronizes sampling clocks of each sensor through network time protocol, eliminating time offset of heterogeneous data. A noise filtering processing module removes noise from original data based on wavelet transform algorithm, retaining irradiance mutation characteristics and temperature fluctuation trend information. The processed dynamic feature vector is stored in a time series database and pushed to a local control algorithm through a message queue. The adaptive droop control algorithm is run in real time, dynamically adjusts inverter power output limit value according to second-level power generation power deviation prediction data and real-time energy storage state of charge value, generates control instructions stored in a control instruction database, and issues the control instructions to an inverter drive circuit through a field bus, realizing millisecond-level response.

[0139] The cloud scheduling center is deployed based on a containerization architecture, receives the hour-level power generation prediction curve and the energy storage efficiency decay prediction result uploaded by the edge computing node. The frequency modulation capacity allocation model integrates the power grid frequency modulation demand curve, the energy storage leasing agreement time period constraint and the battery cluster health state data, and calculates the frequency modulation task allocation weight of each energy storage cluster by using a mixed integer linear programming algorithm. The calculation result generates a clustering frequency modulation capacity allocation proportion table according to the time granularity, is converted into a standardized instruction format by a data packaging module, and is issued to the hour-level control loop through a redundant communication network as a benchmark parameter for the energy storage system to execute the frequency modulation task, and guides the charging and discharging strategy in the next few hours.

[0140] The redundant communication network adopts a heterogeneous architecture of a double CAN bus and a fiber ring network, the main link transmits the second-level inverter power correction instruction to the edge node through the CAN bus, the standby link is activated when it is detected that the main link error rate exceeds a threshold value, and the minute-level charging and discharging power allocation table and the hour-level clustering frequency modulation capacity allocation proportion table are transmitted through the fiber ring network. The communication management module monitors the link state and transmission delay in real time, and the dynamic feature vector processed by the edge node and the inverter actual output power feedback data are compressed and encrypted and then returned to the cloud historical database through the fiber ring network, for online calibration of model parameters and long-term operation analysis.

[0141] The edge computing node performs local data preprocessing and real-time control algorithm to reduce the influence of cloud communication delay on second-level response; the cloud scheduling center integrates long-term prediction data and global constraint conditions to generate an optimized scheduling target; the redundant communication network ensures the reliability of instruction transmission through heterogeneous links. The optimized instructions stored in the edge node maintain control continuity through local caching during communication interruption, and are transmitted to the energy storage system in timestamp order after recovery. The actual output power feedback data is calibrated by the noise filtering threshold of the dynamic feature extraction module through the closed loop link, which improves the accuracy of subsequent data preprocessing.

[0142] The data interaction and instruction coordination logic between components are as follows: the sensor data is processed by the edge node to drive the local control loop to generate second-level instructions, and is uploaded to the cloud to support global scheduling; the long-term scheduling target generated by the cloud is issued to the edge side through the communication network to constrain the minute-level and hour-level optimization model; the feedback data flows bidirectionally between the edge and the cloud, forming a closed loop link from real-time control to long-term optimization. Through hierarchical processing and redundant communication mechanism, the second-level response accuracy and the hour-level scheduling efficiency are synergistically optimized.

[0143] Specifically, the photovoltaic system provided by the application, the redundant communication network comprises:

[0144] The main link error rate is collected, and when the main link error rate exceeds the preset threshold of the double CAN bus, the second-level inverter power correction instruction generated by the synergistic optimization control module is switched to the standby fiber link for transmission.

[0145] During the communication interruption, the unsent optimization instructions generated by the collaborative optimization control module are stored locally, and are transmitted in timestamp order after the communication is restored to the energy storage system;

[0146] The real-time monitored communication delay parameter is input into the minute-level control loop of the collaborative optimization control module, and the optimization period length of the minute-level control loop is dynamically adjusted.

[0147] In the photovoltaic system, the redundant communication network cooperatively guarantees the reliability of instruction transmission and the stability of control through multiple mechanisms. The technical steps are described in detail as follows:

[0148] The main link error rate monitoring module collects the error rate data of the double CAN bus in real time, and uses a sliding window statistical calculation to calculate the average error rate in a unit time. When it is detected that the error rate exceeds the preset threshold, the link switching protocol is triggered, the data transceiver channel of the main link is closed, and the physical layer interface of the standby optical fiber link is activated. The second-level inverter power correction instruction is converted from the CAN bus format to the optical fiber communication protocol format through the protocol conversion module, and is sent to the energy storage system through the optical module. In the switching process, the preamble synchronization mechanism is used to align the transmission timing, so as to avoid the disorder of the timing of the instructions caused by the link switching, and to guarantee the real-time and integrity of the second-level control instructions.

[0149] During the communication interruption, the local buffer area of the edge computing node starts the ring buffer management mechanism, and stores the unsent optimization instructions generated by the collaborative optimization control module in timestamp order. The timestamp is generated by a high-precision clock chip, and is accurate to the millisecond level, so as to realize the traceability and timing consistency of the instruction sequence. After the communication is restored, the transmission module re-packages the data packets according to the timestamp order of the buffered instructions, and preferentially transmits the minute-level charging and discharging instructions with high time effectiveness through the priority queue scheduling algorithm, so as to avoid the control delay caused by the accumulation of instructions, and to maintain the continuity of the charging and discharging strategy of the energy storage system.

[0150] The real-time communication delay monitoring module calculates the end-to-end delay of the instructions from generation to reception through timestamp comparison, and the delay parameter is input into the optimization period adjustment unit of the minute-level control loop after being processed by a smoothing filter. The adjustment unit dynamically expands or compresses the calculation time window of the rolling horizon optimization algorithm based on the difference between the delay parameter and the preset response time threshold. When it is detected that the communication delay is continuously increasing, the iteration period of the optimization algorithm is extended, and the update frequency of the control instructions is reduced. When the delay falls within the threshold, the original period is restored, so as to balance the communication delay accumulation and the real-time demand of control, and to prevent the control instability caused by the delay fluctuation.

[0151] The logical association between each mechanism is as follows: the main link switching mechanism ensures the reliable transmission of key instructions through real-time bit error rate monitoring, avoiding interference of communication quality degradation on second-level control; the local cache and retransmission mechanism maintains instruction integrity when communication is interrupted, and schedules retransmission data according to time priority after recovery, supporting control continuity and timing consistency; the delay parameter is fed back to the control loop to dynamically adjust and optimize the period, realizing the real-time coupling of communication state and control parameter. The three constitute a "link redundancy-data integrity-control stability" collaborative architecture, which improves the robustness of the system under complex working conditions through multi-level linkage of communication quality monitoring, instruction storage logic and control period adaptation.

[0152] In the technical solution of the redundant communication network, the main link switching mechanism solves the problem of physical layer transmission reliability, the cache and retransmission mechanism deals with network layer interruption scenarios, and the delay feedback mechanism optimizes the application layer control logic. The transmission reliability of second-level instructions is realized through the dual-link heterogeneous architecture and protocol conversion, the integrity of minute-level instructions depends on cache management and priority scheduling, and the stability of hour-level scheduling instructions is guaranteed by dynamic calibration of delay parameters. Each level of mechanism is interconnected through timestamp alignment, standardized data encapsulation format and parameter feedback interface, forming a closed-loop communication optimization system to meet the multiple requirements of real-time, reliability and self-adaptation of industrial process control systems.

[0153] Specifically, the photovoltaic system described in the application, the cloud scheduling center comprises:

[0154] Receiving the hour-level power generation prediction curve generated by the prediction model module in the photovoltaic assembly and the energy storage health state data fed back by the energy storage control unit in the photovoltaic assembly;

[0155] According to the preset power grid frequency modulation demand and the preset energy storage leasing agreement constraint, the frequency modulation task allocation weight of each energy storage cluster is calculated;

[0156] The clustering frequency modulation capacity allocation proportion table is sent to the hour-level control loop of the photovoltaic assembly as an input constraint condition of the mixed integer linear programming model of the hour-level control loop.

[0157] In the photovoltaic system described in the application, the cloud scheduling center realizes global frequency modulation capacity allocation through data fusion and optimization model. The following describes each technical step in detail:

[0158] The data receiving module of the cloud scheduling center subscribes to the hourly power generation prediction curve published by the photovoltaic component prediction model module through the message middleware, synchronously polls the battery cluster health state data fed back by the energy storage control unit, including the cycle number, internal resistance change rate and capacity attenuation gradient. The prediction curve and the health state data are unified into a standardized time series data model by the format conversion module and stored in a distributed database. The data verification unit verifies the timestamp consistency and numerical validity, and eliminates abnormal data samples to provide high confidence input for frequency modulation task allocation.

[0159] The frequency modulation task allocation algorithm is designed based on a multi-objective optimization framework. The objective function integrates the capacity gap of the grid frequency modulation demand curve and the service period constraint agreed in the energy storage leasing agreement. The battery cluster health state score is used as the life loss weight coefficient, and the frequency modulation task priority of each cluster is calculated by weighted summation method. The algorithm preferentially allocates the battery clusters with high health score and sufficient remaining service period to execute high-frequency frequency modulation tasks, generates a clustering frequency modulation capacity allocation proportion table, and clearly defines the power output upper limit and frequency modulation response weight of each cluster in different periods.

[0160] The clustering frequency modulation capacity allocation proportion table is converted into a constraint condition format that can be analyzed by the mixed integer linear programming model by the protocol packaging module, including the power output boundary, period marker and life loss coefficient. After adding version markers and timestamps to the constraint condition data package, it is sent to the model input interface of the hourly control loop through the redundant communication network. The version comparison mechanism checks the timeliness of the data package at the receiving end to avoid conflicts between old and new constraint conditions due to network delay, and maintains the time sequence consistency of the scheduling instructions.

[0161] The collaborative logic of the data receiving module and the frequency modulation task allocation algorithm is as follows: the standardized data model provides a unified input format for multi-objective optimization, eliminating compatibility problems of heterogeneous data sources; the weighted scoring model balances grid demand and device life, preventing systemic risks caused by local overload; the constraint condition packaging mechanism converts the optimization results into boundary parameters recognizable by the control model, realizing precise mapping from global scheduling target to local control instruction. During the delivery process, the timestamp alignment mechanism ensures that the proportion table and the real-time grid demand are updated synchronously, and the version marker prevents instruction overlap.

[0162] The data interaction between the cloud scheduling center and the hourly control loop is realized through standardized interfaces. The prediction curve and health data drive the frequency modulation task allocation algorithm to generate a proportion table, which is used as a hard constraint condition for mixed integer linear programming to limit the solution space of the optimization model. The frequency modulation instruction sequence output by the hourly control loop is uploaded to the cloud historical database through the feedback channel for iterative calibration of model parameters. This closed-loop mechanism continuously optimizes the scheduling strategy through bidirectional data flow, improving the accuracy and economy of long-term frequency modulation capacity allocation.

[0163] Specifically, the photovoltaic system also includes a fault handling module, which is configured to:

[0164] When the sensor group of the photovoltaic module fails, the historical synchronous data stored by the dynamic feature extraction module is called to generate a virtual sensor value in combination with the irradiance spatial mapping model of the adjacent photovoltaic array;

[0165] When the interruption duration of the redundant communication network exceeds a set threshold, the energy storage system is switched to an off-grid operation mode, and non-critical loads are cut off according to the frequency regulation task allocation weight calculated by the cloud scheduling center;

[0166] When the power device of the energy storage control unit in the photovoltaic module is detected to have overcurrent, the charging and discharging rate is reduced, and the junction temperature prediction model is adjusted by the temperature rise state correction value output by the prediction model module to drive the switching frequency.

[0167] In the photovoltaic system, the cloud scheduling center realizes global frequency regulation capacity allocation through data fusion and optimization models. The technical steps are described in detail as follows:

[0168] The data receiving module of the cloud scheduling center subscribes to the hour-level power generation prediction curve published by the photovoltaic module prediction model module through the message middleware, and synchronously polls the battery cluster health state data fed back by the energy storage control unit, including the cycle number, internal resistance change rate and capacity attenuation gradient. The prediction curve and the health state data are uniformly converted into a standardized time series data model by a format conversion module and stored in a distributed database. The data verification unit verifies the timestamp consistency and numerical validity, and eliminates abnormal data samples to provide high confidence input for frequency regulation task allocation.

[0169] The frequency regulation task allocation algorithm is designed based on a multi-objective optimization framework, and the target function integrates the capacity gap of the grid frequency regulation demand curve and the service period constraint agreed in the energy storage leasing agreement. The battery cluster health state score is used as a life loss weight coefficient, and the frequency regulation task priority of each cluster is calculated by a weighted summation method. The algorithm preferentially allocates the battery clusters with high health scores and sufficient remaining service periods to execute high-frequency frequency regulation tasks, generates a clustering frequency regulation capacity allocation proportion table, and clearly defines the power output upper limit and frequency regulation response weight of each cluster in different periods.

[0170] The clustering frequency regulation capacity allocation proportion table is converted into a constraint condition format that can be analyzed by a mixed integer linear programming model by a protocol packaging module, including power output boundaries, period markers and life loss coefficients. After adding version markers and timestamps to the constraint condition data package, it is sent to the model input interface of the hour-level control loop through the redundant communication network. The version comparison mechanism checks the timeliness of the data package at the receiving end to avoid conflicts between new and old constraint conditions due to network delay, and maintains the time sequence consistency of the scheduling instructions.

[0171] The cooperative logic of the data receiving module and the frequency modulation task allocation algorithm is as follows: the standardized data model provides a unified input format for multi-objective optimization, eliminating the compatibility problem of heterogeneous data sources; the weighted scoring model balances the power grid demand and the equipment life, preventing systemic risks caused by local overload; the constraint condition packaging mechanism converts the optimization results into boundary parameters recognizable by the control model, realizing accurate mapping from global scheduling targets to local control instructions. In the issuing process, the timestamp alignment mechanism ensures that the proportion table and the real-time power grid demand are updated synchronously, and the version marker prevents instruction overlap execution.

[0172] The data interaction between the cloud scheduling center and the hour-level control loop is realized through a standardized interface. The prediction curve and health data drive the frequency modulation task allocation algorithm to generate a proportion table, which is used as a hard constraint condition for the mixed integer linear programming, limiting the solution space of the optimization model. The frequency modulation instruction sequence output by the hour-level control loop is uploaded to the cloud historical database through the feedback channel for iterative calibration of model parameters. This closed-loop mechanism continuously optimizes the scheduling strategy through bidirectional data flow, improving the accuracy and economy of long-term frequency modulation capacity allocation.

[0173] Specifically, the photovoltaic system described in the present application, the fault handling module is further configured to: after the communication is restored, compare the time difference between the locally cached unsent optimization instructions and the latest instructions issued by the cloud scheduling center, and update the control instruction database of the edge computing node using a preset conflict resolution algorithm;

[0174] According to the deviation between the state of charge real-time data returned through the feedback channel and the virtual estimated value generated based on the constant current and constant voltage charging and discharging sequence, the charging and discharging sequence is restarted to calibrate the energy storage capacity;

[0175] The fault event record is input into the case library, and a typical fault response strategy template is generated through cluster analysis, which is used to optimize the adaptive fault tolerance parameters of the cooperative optimization control module.

[0176] In the photovoltaic system described in the present application, the fault handling module realizes dynamic optimization of fault tolerance strategies through a closed-loop feedback and self-learning mechanism. The following describes the technical steps in detail:

[0177] After the communication is restored, the time difference comparison module parses the locally cached unsent optimization instructions and the latest instruction metadata issued by the cloud scheduling center, extracts the instruction generation timestamp and version marker. The sliding window matching algorithm is used to align the time sequence, and the conflict resolution rules are based on priority and timeliness weight: the cloud instruction version updates the local cache instruction; the local instruction contains real-time optimization results that have not been uploaded, triggering data merging logic. The updated instruction set is written into the control instruction database of the edge computing node through the database transaction mechanism, maintaining data consistency and transaction integrity, and avoiding control conflicts caused by instruction overlap.

[0178] The constant current and constant voltage charge-discharge calibration program calculates the absolute deviation of the state of charge real-time data returned by the feedback channel from the virtual estimated value. The virtual estimated value is generated based on the segmented charge-discharge sequence of the rated current charge-discharge to the voltage threshold in the constant current stage and the constant voltage maintenance to the current cutoff value in the constant voltage stage. The battery terminal voltage and current integral data are synchronously collected during the calibration process, and the nominal capacity parameters of the energy storage system are dynamically corrected through the capacity calculation model. The correction coefficient is combined with the battery surface temperature data for weighted adjustment, the health state score of the battery cluster is updated and fed back to the prediction model module, and the capacity evaluation accuracy is improved.

[0179] The fault event record module encapsulates the key parameters (such as fault type, timestamp, and impact range) of communication interruption, sensor failure, and power device overcurrent event into a structured log and stores it in a distributed case library. The clustering analysis engine uses a density-based noise application spatial clustering algorithm to identify the spatiotemporal distribution characteristics and relevance of fault events, and extracts high-frequency fault patterns to generate response strategy templates. The template includes fault trigger conditions, fault tolerance parameter adjustment rules, and control instruction correction suggestions, which update the adaptive fault tolerance parameters of the collaborative optimization control module through the parameter configuration interface, and optimize the droop coefficient mapping relationship and the rolling time domain window length.

[0180] The logical association between the mechanisms is as follows: the time scale difference comparison and conflict resolution realize the integrity and timeliness of the control instructions after the communication is restored, and prevent control conflicts caused by data inconsistency; the charge-discharge calibration program dynamically corrects the capacity parameters based on the measured data to improve the state evaluation accuracy of the energy storage system; the clustering analysis and strategy template generation of the fault case library drive the adaptive optimization of the control module parameters. The three form a closed-loop fault tolerance link of "data synchronization-state calibration-strategy iteration", which enhances the adaptability of the system to complex fault scenarios through the synergistic effect of historical experience and real-time feedback.

[0181] In the technical scheme of the application, the instruction synchronization mechanism solves the data consistency problem after the communication interruption is restored, the calibration program deals with the battery capacity evaluation error, the case library self-learning mechanism optimizes the long-term fault tolerance strategy, the historical database is the basic storage layer of system operation data, focusing on raw data management and recovery; the case library is the experience induction layer of fault knowledge, focusing on strategy iteration and active fault tolerance. The conflict resolution rules and the database transaction mechanism guarantee the atomicity of instruction updating; the segmented charge-discharge calibration combined with temperature weighted correction improves the capacity calibration accuracy; the fault mode template extracted by the clustering algorithm provides an empirical basis for parameter optimization. Through multi-level strategy iteration and dynamic parameter adjustment, the scheme realizes the upgrade of fault tolerance control from passive response to active prevention, meeting the requirements of industrial control systems for reliability and adaptability.

[0182] The technical features of the application are explained as follows:

[0183] Sensor group:

[0184] Front irradiance sensor: A multispectral probe is installed vertically on the front glass surface of the bifacial module, covering the spectral range from visible light to near-infrared. A photodiode array is used to capture real-time direct light intensity and spectral distribution data.

[0185] Back reflectivity sensor: An integrated wide-angle lens and polarizing filter are arranged obliquely on the edge of the module backboard to measure the multi-angle incident intensity of water surface reflected light. The dynamic reflection coefficient is output by combining the reflectivity calculation model.

[0186] Temperature sensor group: Platinum resistance PT1000 probes are selected and distributed in a grid along the diagonal and center point of the photovoltaic array. Four temperature measurement nodes are arranged per square meter. Four-wire measurement method is used to eliminate wire resistance errors. Horizontal and vertical temperature gradient data of the module surface are collected.

[0187] Monitoring unit: High-precision coulomb meter and Hall current sensor are embedded in the bus cabinet of the energy storage system battery cluster. The state of charge (SOC) distribution and charge / discharge rate of the battery cluster are collected non-contactly.

[0188] Dynamic feature extraction module:

[0189] Timestamp alignment unit: Network time protocol (NTP) is used to synchronize the sampling clocks of heterogeneous sensors. Millisecond-level timestamps are added to irradiance data through RS-485 bus. Temperature data is aligned using hardware interrupts through SPI interface. CAN bus protocol has a timestamp field to synchronize energy storage state data, eliminating time sequence deviations of multi-source data.

[0190] Noise filtering processing module: Wavelet transform algorithm is used to separate high-frequency mutation features of irradiance data (such as cloud cover). Moving average algorithm is used to smooth the baseline drift of temperature data. Sliding window mean filtering is used to suppress charge / discharge rate pulse interference, retaining valid information.

[0191] Dynamic feature vector: The irradiance difference coefficient is calculated by the front and back light intensity ratio. The temperature fluctuation variance is quantified by the standard deviation of the grid temperature measurement nodes. The maximum difference of SOC between battery clusters is used to evaluate the dispersion of the state of charge. The fusion forms a standardized multi-dimensional vector.

[0192] Prediction model module:

[0193] Long short-term memory network (LSTM): The double-channel input layer receives front and back irradiance time series data. The direct light channel uses a 5-second sliding window to extract second-level fluctuation features. The reflected light channel eliminates random noise by time decay weight. The output is a 5-second power deviation prediction sequence.

[0194] Kalman filter: Based on the component heat conduction model, the state transition equation is constructed, the temperature gradient data is input, the temperature rise rate and heat dissipation efficiency parameters are dynamically corrected, and the temperature rise state correction value of the internal hot spot area of the component is output.

[0195] Energy storage efficiency attenuation model: According to the historical correlation between state of charge and charge / discharge rate, the capacity attenuation gradient parameter is updated online through incremental learning mechanism to generate efficiency attenuation prediction curve as charge / discharge depth constraint condition.

[0196] Cooperative optimization control module:

[0197] Second-level control loop: Adopting adaptive droop control algorithm, dynamically adjusting droop coefficient according to power generation power deviation prediction value and real-time state of charge, preferentially calling battery cluster with low state of charge dispersion to execute power compensation, generating inverter pulse width modulation instruction.

[0198] Minute-level control loop: Based on rolling horizon optimization algorithm, taking temperature rise correction value as thermodynamic constraint and efficiency attenuation prediction as life loss constraint, solving the optimal matching path of charge / discharge rate and maximum power point tracking voltage, generating charge / discharge power allocation table.

[0199] Hour-level control loop: Integrating grid frequency modulation period demand and energy storage leasing agreement, constructing mixed integer linear programming model, taking maximum frequency modulation income as target, generating cluster frequency modulation capacity allocation proportion table, distributing frequency modulation task weight according to time granularity.

[0200] Instruction execution module:

[0201] Inverter interface unit: converting digital instructions into pulse width modulation signals to drive IGBT switching devices to regulate DC bus voltage, synchronously collecting output voltage / current waveform, calculating actual power value.

[0202] Energy storage control unit: analyzing charge / discharge power allocation table, dynamically adjusting battery cluster switching sequence through multi-channel relay array, combining coulomb counting method to monitor charge / discharge efficiency in real time.

[0203] Feedback channel: packaging actual power waveform and efficiency data into standardized data packets, returning to dynamic feature extraction module, dynamically updating wavelet denoising threshold based on statistical features such as harmonic distortion rate.

[0204] Redundant communication network:

[0205] Dual CAN bus and fiber ring network: main link transmits second-level instructions through CAN bus, standby fiber link is activated when bit error rate exceeds standard, ensuring instruction transmission reliability; local cache instructions when communication is interrupted, and retransmit according to time stamp order after recovery.

[0206] Communication delay dynamic adjustment: Real-time monitoring of end-to-end delay, dynamically expanding or compressing the rolling time domain window of the minute-level control loop, balancing delay accumulation and control real-time requirements.

[0207] Cloud dispatch center:

[0208] Frequency regulation task allocation algorithm: Integrating grid frequency regulation demand curve and energy storage health status data, weighted calculation of battery cluster frequency regulation priority, generation of clustering frequency regulation capacity allocation proportion table, encapsulation as constraint conditions of mixed integer linear programming.

[0209] Data verification and version management: Verify timestamp consistency and exclude abnormal data; add version markers when issuing instructions to avoid conflicts between new and old constraint conditions.

[0210] Fault handling module:

[0211] Virtual sensor value generation: Call historical contemporaneous data and adjacent array irradiance spatial mapping model to generate alternative data for failed sensors, maintaining control continuity.

[0212] Off-grid operation and load shedding: Switch to off-grid mode when communication interruption exceeds threshold, shed non-critical loads according to frequency regulation task weight, and ensure core load power supply.

[0213] Overcurrent protection and junction temperature estimation: Detect power device overcurrent, dynamically adjust IGBT switching frequency combined with temperature rise correction value, and suppress thermal stress accumulation.

[0214] Case library and strategy template: Extract high-frequency fault patterns through clustering analysis, generate fault-tolerant parameter adjustment rules, and optimize droop coefficient mapping and rolling time domain window length.

[0215] Long short-term memory network (LSTM) model: Used to model second-level irradiance changes of double-sided photovoltaic arrays and predict power generation power deviation within the next 5 seconds.

[0216] Implementation:

[0217] Dual-channel input layer: Receive time series data of front irradiance and back reflectance intensity respectively. The direct light channel uses a 5-second sliding window to extract irradiance mutation features (such as instantaneous fluctuations caused by cloud cover), and the reflectance channel calculates dynamic reflectance contribution values through a time decay weight algorithm to eliminate water surface reflection noise interference.

[0218] Gated recurrent unit (GRU): Fuse dual-channel outputs to generate power generation power deviation prediction sequences, align timestamps with real-time power sampling values, and form second-level control loop deviation correction benchmarks.

[0219] Implementation:

[0220] Kalman filter: assess the temperature rise state of the photovoltaic module, correct the heat dissipation efficiency parameter.

[0221] State transition equation: based on the heat conduction model of the module, define the temperature rise rate and heat dissipation efficiency as state variables, and the surface temperature gradient as the observation variable.

[0222] Noise processing: the process noise matrix is dynamically adjusted according to the environmental temperature fluctuation covariance, and the observation noise matrix is calibrated by the measurement accuracy of the temperature sensor.

[0223] Output correction value: through the iterative prediction and update steps, output the temperature rise state correction value of the internal hot spot area of the module, and drive the heat dissipation strategy optimization.

[0224] Energy storage efficiency decay model: predict the capacity attenuation trend of the energy storage system, and optimize the charge and discharge depth constraint.

[0225] Implementation:

[0226] Incremental learning mechanism: according to the historical correlation between state of charge (SOC) and charge and discharge rate, update the model parameters online, and calibrate the effective data samples through sliding window filtering in the process, and eliminate the interference of abnormal events such as overvoltage protection.

[0227] Output result: generate efficiency decay prediction curve as the charge and discharge power allocation constraint condition of the minute-level control loop, limit the charge and discharge depth of the high-loss battery cluster.

[0228] Adaptive droop control algorithm: generate inverter power correction instructions in the second-level control loop to suppress grid frequency fluctuations.

[0229] Implementation:

[0230] Droop coefficient dynamic mapping: establish the correlation between the droop coefficient and the state of charge (SOC) of the energy storage system, and adjust the inverter output limit value in real time when the power deviation caused by irradiance mutation is detected.

[0231] Priority calling strategy: preferentially schedule the battery cluster with low state of charge dispersion to execute power compensation, and drive the inverter power switching device through pulse width modulation signal.

[0232] Rolling horizon optimization algorithm (RTO): optimize the matching relationship between energy storage charge and discharge rate and maximum power point tracking voltage in the minute-level control loop.

[0233] Implementation:

[0234] Multi-stage decision model: divide the time window into discrete state nodes, and at each node, take the temperature rise correction value as the thermodynamic constraint and the efficiency decay prediction as the life loss constraint to solve the multi-objective optimization charge and discharge path.

[0235] Dynamic programming solution: generate charge-discharge power allocation table, transmit to energy storage converter through CAN bus, adjust battery cluster switching order and equalization strategy.

[0236] Mixed integer linear programming (MILP): generate cluster frequency modulation capacity allocation ratio table in hour-level control loop to meet grid frequency modulation demand.

[0237] Implementation:

[0238] Objective function: maximize frequency modulation income as the core, integrate energy storage lease agreement period constraints, battery life loss coefficient and remaining capacity restrictions.

[0239] Constraint condition: define frequency modulation capacity, service period, and charge-discharge cycle number as decision variables, and generate frequency modulation task weight table divided by time granularity.

[0240] Redundant communication network architecture: ensure real-time and reliability of instruction transmission.

[0241] Implementation:

[0242] Dual CAN bus and fiber ring network: main link transmits second-level instructions, backup link activates when error rate exceeds standard; local cache instructions during communication interruption, and retransmit in timestamp order after recovery.

[0243] Delay dynamic adjustment: real-time monitoring of communication delay parameters, dynamically expanding or compressing the optimization period of minute-level control loop, balancing delay accumulation and control real-time demand.

[0244] Case library and clustering analysis (DBSCAN): generate typical fault response strategy template, optimize fault tolerance parameters.

[0245] Implementation:

[0246] Fault event structured record: store fault type, timestamp, impact range and other parameters, identify high-frequency fault patterns through density-based spatial clustering (DBSCAN).

[0247] Strategy template generation: extract fault trigger conditions and processing rules, optimize adaptive droop coefficient mapping relationship and rolling time domain window length.

[0248] The embodiment of the application is based on the collaborative control demand of the double-sided photovoltaic array and the rental energy storage system in the fish-light complementary scene, and realizes the dynamic matching of multiple time scales through the following steps: the front surface irradiance sensor in the sensor group is vertically installed on the front surface glass layer of the component by using a multi-spectral probe, covers the visible light to near-infrared band, and captures the direct light intensity and spectral distribution in real time; the back reflectivity sensor is integrated with a wide-angle lens and a polarization filter, and is arranged obliquely at the edge of the back plate to measure the multi-angle incident intensity of the water surface reflected light, and the light intensity data is transmitted to the dynamic characteristic extraction module through the RS-485 bus. The temperature sensor group selects a platinum resistance PT1000 probe, is distributed in a grid shape along the diagonal and the center point of the photovoltaic array, four temperature measurement nodes are arranged per square meter, the four-wire measurement is used to eliminate the resistance error of the wire, and the temperature gradient data is transmitted through the SPI interface. The energy storage monitoring unit is configured with a high-precision coulomb meter and a Hall current sensor, and the state of charge and the charging and discharging rate of the battery cluster are collected in a non-contact manner, and the data is transmitted to the dynamic characteristic extraction module through the CAN bus protocol. The dynamic characteristic extraction module synchronizes the heterogeneous data time stamp based on the network time protocol, uses the wavelet transform and the moving average combination algorithm to retain the high-frequency mutation characteristics of the irradiance data, filter the baseline drift of the temperature data, and smooth the pulse interference of the charging and discharging rate data, and generates a dynamic characteristic vector containing the irradiation difference coefficient, the temperature fluctuation variance and the state of charge dispersion. In the prediction model module, the long short-term memory network double-channel input layer receives the front and back surface irradiation time series data respectively, the direct light channel adopts a 5-second sliding window to extract the second-level fluctuation characteristics, the reflected light channel calculates the dynamic reflectivity contribution value through the time attenuation weight, and outputs the future 5-second power generation power deviation prediction sequence; the Kalman filter constructs a component heat conduction state space model based on the temperature gradient data, iteratively corrects the temperature rise rate and the heat dissipation efficiency parameters, and outputs the temperature rise state correction value; the energy storage efficiency attenuation model updates the capacity attenuation gradient parameters online according to the historical correlation of the state of charge and the charging and discharging rate through an incremental learning mechanism, and generates an efficiency attenuation prediction curve. In the collaborative optimization control module, the second-level control ring adopts an adaptive droop control algorithm, dynamically adjusts the inverter power output limit value, and preferentially calls the battery cluster with low state of charge dispersion to execute power compensation; the minute-level control ring is based on a rolling horizon optimization algorithm, takes the temperature rise correction value and the efficiency attenuation prediction as constraint conditions, and solves the optimal matching path of the charging and discharging rate and the maximum power point tracking voltage; the hour-level control ring integrates the grid frequency modulation time period demand and the energy storage rental agreement through a mixed integer linear programming model, generates a clustering frequency modulation capacity distribution proportion table, and distributes the frequency modulation task weight according to the time granularity. The instruction execution module converts the inverter power correction instruction into a pulse width modulation signal to drive the IGBT switching device, and synchronously collects the actual output power waveform; the energy storage control unit analyzes the charging and discharging power distribution table, dynamically adjusts the battery cluster switching sequence through a multi-channel relay array; the feedback channel returns the actual power and efficiency data to the dynamic characteristic extraction module, and dynamically updates the wavelet denoising threshold based on the harmonic distortion rate.The redundant communication network adopts a double CAN bus and a fiber ring network architecture, the main link transmits second-level instructions, the standby link is activated when the error code rate exceeds the standard, and the local cache does not send instructions during communication interruption and transmits them in timestamp order. The fault handling module calls historical contemporaneous data and adjacent array irradiation mapping model to generate virtual sensor values, switches to off-grid mode when the communication interruption exceeds the threshold and removes non-critical loads according to the frequency modulation weight, adjusts the switching frequency in combination with the temperature rise correction value when the power device overflows, and generates a strategy template to optimize fault tolerance parameters through cluster analysis of fault event records. The edge computing node performs local data preprocessing and second-level control algorithm, the cloud scheduling center generates a global frequency modulation target based on the hour-level prediction curve and health status data, and realizes multi-level cooperation of second-level response, minute-level optimization and hour-level scheduling through closed-loop data flow, solving the cooperative control problem of dynamic fluctuation of double-sided photovoltaic output and response lag of energy storage.

[0249] The technical scheme of the present application solves the technical problems in the background art by using the following technical scheme:

[0250] The sensor group collects real-time data of the front irradiance, back reflectivity and temperature gradient of the double-sided photovoltaic array, and synchronously obtains the state of charge and charge / discharge rate of the energy storage system. The dynamic characteristic extraction module generates a dynamic characteristic vector containing irradiance difference coefficient, temperature fluctuation variance and state of charge dispersion through timestamp alignment and noise filtering processing. This vector is used as the input of the prediction model module to drive the long short-term memory network (LSTM) to predict the second-level power generation power deviation, the Kalman filter to evaluate the component temperature rise state, and the energy storage efficiency decay model to quantify the battery performance degradation trend, providing multi-dimensional dynamic benchmarks for hierarchical control.

[0251] The cooperative optimization control module is divided into second-level, minute-level and hour-level control loops. The second-level control loop is based on an adaptive droop control algorithm, dynamically adjusts the inverter power output limit value according to the power generation power deviation prediction value and real-time state of charge, and realizes millisecond-level power grid frequency fluctuation suppression; the minute-level control loop adopts a rolling horizon optimization algorithm, takes the temperature rise correction value and efficiency decay prediction as constraint conditions, and solves the optimal matching path of energy storage charge / discharge rate and maximum power point tracking voltage; the hour-level control loop integrates the power grid frequency modulation demand and energy storage leasing agreement through a mixed integer linear programming model, generates a clustering frequency modulation capacity distribution proportion table, and organizes the frequency modulation task weight according to the time granularity. The three control loops are interconnected through a data bus to form a closed-loop instruction cooperation from transient response to long-term scheduling.

[0252] Redundant communication network adopts dual CAN bus and fiber ring network architecture, the main link transmits second-level instructions, the standby link is activated when the bit error rate exceeds the standard, which guarantees the reliability of instruction transmission; during communication interruption, local cache does not send instructions, and after recovery, it is transmitted in timestamp order. The fault handling module generates virtual sensor values, removes off-grid load, and adjusts the junction temperature estimation strategy to deal with sensor failure, communication interruption and power device overcurrent scenarios. The feedback channel returns the actual output power and efficiency data to the dynamic feature extraction module, which dynamically calibrates the noise filtering threshold; the case library generates typical fault response strategy templates through cluster analysis, optimizes adaptive fault tolerance parameters, and improves the system's self-learning ability. Edge computing nodes and cloud scheduling centers cooperate to cover second-level control and hour-level resource allocation, ultimately solving the multi-time scale matching problem of photovoltaic output dynamic fluctuation and energy storage response delay.

Claims

1. A photovoltaic module, characterized in that, include: The sensor array is used to collect data on the front irradiance, back reflectance, and ambient temperature of the bifacial photovoltaic array, and simultaneously acquire the state of charge and charge / discharge rate of the energy storage system. The dynamic feature extraction module performs timestamp alignment and noise filtering on the collected data to generate dynamic feature vectors including irradiance differences, temperature fluctuations, and energy storage state parameters. The prediction model module receives front irradiance, back reflectance, and temperature fluctuation data from the dynamic feature vector. It uses a preset long short-term memory network to model the second-level irradiance change of the front irradiance, generating second-level power generation deviation prediction data. At the same time, based on the temperature fluctuation data, it inputs a preset Kalman filter to evaluate the component temperature rise status, outputs the temperature rise status correction value, and updates the preset energy storage efficiency degradation model parameters according to the state of charge and charge / discharge rate data, generating energy storage efficiency degradation prediction results. The collaborative optimization control module includes a second-level control loop, a minute-level control loop, and an hour-level control loop. The second-level power generation deviation prediction data is input into the preset second-level control loop, and an inverter power correction command is generated through a preset adaptive droop control algorithm. The temperature rise state correction value and energy storage efficiency decay prediction result are input into the minute-level control loop, and an energy storage charging and discharging power allocation table is generated by combining a preset rolling time-domain optimization algorithm. The grid frequency regulation period constraint is input into the hour-level control loop, and a clustered frequency regulation capacity allocation ratio table is generated through a preset mixed integer linear programming. The instruction execution module receives the inverter power correction instruction, the energy storage charging and discharging power allocation table, and the cluster frequency modulation capacity allocation ratio table. It adjusts the maximum power point tracking voltage parameters of the bifacial photovoltaic array and the charging and discharging rate of the energy storage system. The adjusted parameters of the bifacial photovoltaic array and the actual output power data are transmitted to the dynamic feature extraction module through the feedback channel to update the noise filtering threshold of the dynamic feature vector.

2. The photovoltaic module according to claim 1, characterized in that, The sensor group includes: A front irradiance sensor and a back reflectance sensor are respectively deployed on the front and back surfaces of the double-sided component to collect the intensity of direct and reflected light, and transmit the collected direct and reflected light intensity data to the dynamic feature extraction module through a data bus. A temperature sensor array, distributed along the edge and center of the photovoltaic array, is used to measure the temperature gradient on the surface of the component and input the temperature gradient data into the dynamic feature extraction module via a bus protocol. The monitoring unit is connected to the battery clusters of the energy storage system and is used to acquire the state of charge distribution and charge / discharge rate of each cluster. The state of charge distribution and charge / discharge rate data are then input into the dynamic feature extraction module through the communication interface.

3. The photovoltaic module according to claim 2, characterized in that, The prediction model module includes: The direct and reflected light intensity data collected by the sensor group are input into the dual-channel input layer of a preset long short-term memory network to generate second-level power generation deviation prediction data. The temperature gradient data collected by the temperature sensor group is input into a preset Kalman filter to construct the state transition equation of component temperature rise and heat dissipation efficiency, and the temperature rise state correction value is output to the collaborative optimization control module. Based on the state of charge distribution and charge / discharge rate data collected by the monitoring unit, the parameters of the preset energy storage efficiency degradation model are updated to generate energy storage efficiency degradation prediction results.

4. The photovoltaic module according to claim 3, characterized in that, The collaborative optimization control module includes: The second-level control loop receives the second-level power generation deviation prediction data generated by the prediction model module and the real-time state of charge data collected by the monitoring unit, and generates inverter power correction commands through a preset adaptive droop control algorithm. The minute-level control loop, based on the temperature rise state correction value output by the prediction model module and the energy storage efficiency decay prediction result, uses a preset dynamic programming algorithm to optimize the matching relationship between the energy storage charge and discharge rate and the maximum power point tracking voltage parameter. The hourly control loop, in conjunction with the grid frequency regulation time constraints, generates a clustered frequency regulation capacity allocation ratio table through a preset mixed integer linear programming, and then sends the clustered frequency regulation capacity allocation ratio table to the energy storage system.

5. The photovoltaic module according to claim 4, characterized in that, The instruction execution module includes: The inverter interface unit converts the inverter power correction command generated by the second-level control loop of the collaborative optimization control module into a pulse width modulation signal and acquires the actual output power waveform. The energy storage control unit adjusts the switching sequence of battery clusters according to the energy storage charging and discharging power allocation table generated by the minute-level control loop of the collaborative optimization control module, and simultaneously monitors the charging and discharging efficiency; The feedback channel transmits the actual output power waveform collected by the inverter interface unit and the charge / discharge efficiency data monitored by the energy storage control unit back to the dynamic feature extraction module, which is used to update the noise filtering threshold of the dynamic feature vector.

6. A photovoltaic system applied to a photovoltaic module as described in any one of claims 1 to 5, characterized in that, include: An edge computing node is deployed on the photovoltaic array side and connected to the sensor group of the photovoltaic module. The edge computing node stores optimization instructions to the control instruction database, receives front irradiance, back reflectance and ambient temperature data collected by the sensor group, performs timestamp alignment and noise filtering processing of the dynamic feature extraction module in the photovoltaic module, and runs the adaptive droop control algorithm of the second-level control loop in the photovoltaic module. The cloud-based dispatch center connects to the edge computing nodes through a redundant communication network, receives the hourly power generation prediction curve and energy storage efficiency decay prediction results generated by the prediction model module in the photovoltaic module, generates a global frequency regulation capacity allocation target based on the preset grid frequency regulation requirements and preset energy storage leasing agreement constraints, and sends the global frequency regulation capacity allocation target to the hourly control loop of the photovoltaic module. The redundant communication network adopts a dual CAN bus and fiber optic ring network architecture to transmit the second-level inverter power correction command, the minute-level energy storage charging and discharging power allocation table, and the hour-level clustered frequency modulation capacity allocation ratio table generated by the collaborative optimization control module in the photovoltaic module to the edge computing node and the energy storage system. At the same time, the dynamic feature vector and actual output power feedback data processed by the edge computing node are uploaded to the cloud scheduling center.

7. The photovoltaic system according to claim 6, characterized in that, The redundant communication network includes: The main link bit error rate is collected. When the main link bit error rate exceeds the preset threshold of the dual CAN bus, the system switches to the backup fiber optic link to transmit the second-level inverter power correction command generated by the collaborative optimization control module. During the communication interruption, the unsent optimization instructions generated by the collaborative optimization control module are stored locally and then retransmitted to the energy storage system in timestamp order after communication is restored. The real-time monitored communication delay parameters are input into the minute-level control loop of the collaborative optimization control module to dynamically adjust the optimization cycle duration of the minute-level control loop.

8. The photovoltaic system according to claim 7, characterized in that, The cloud-based dispatch center includes: Receive the hourly power generation prediction curve generated by the prediction model module in the photovoltaic module and the energy storage health status data fed back by the energy storage control unit in the photovoltaic module; Based on the preset grid frequency regulation requirements and the preset energy storage leasing agreement constraints, the frequency regulation task allocation weight of each energy storage cluster is calculated. The cluster frequency modulation capacity allocation ratio table is sent to the hourly control loop of the photovoltaic module as the input constraint condition of the mixed integer linear programming model of the hourly control loop.

9. The photovoltaic system according to claim 8, characterized in that, It also includes a fault handling module, which is configured as follows: When the sensor group of the photovoltaic module fails, the historical data stored in the dynamic feature extraction module is called and combined with the irradiance spatial mapping model of the adjacent photovoltaic array to generate virtual sensing values. When the interruption duration of the redundant communication network exceeds a set threshold, the energy storage system is switched to off-grid operation mode, and non-critical loads are cut off according to the frequency regulation task allocation weight calculated by the cloud dispatch center. When an overcurrent is detected in the power device of the energy storage control unit in the photovoltaic module, the charging and discharging rate is reduced and the junction temperature prediction model driven by the temperature rise state correction value output by the prediction model module is triggered to adjust the switching frequency.

10. The photovoltaic system according to claim 9, characterized in that, The fault handling module is further configured to: after the communication is restored, compare the timestamp difference between the unsent optimization instructions cached locally and the latest instructions issued by the cloud scheduling center, and update the control instruction database of the edge computing node using a preset conflict resolution algorithm; Based on the deviation between the real-time state of charge data returned by the feedback channel and the virtual estimate generated based on the constant current and constant voltage charge-discharge sequence, the charge-discharge sequence is initiated to recalibrate the energy storage capacity. The fault event records are entered into the case library, and typical fault response strategy templates are generated through cluster analysis, which are used to optimize the adaptive fault tolerance parameters of the collaborative optimization control module.

Images