Control method and system of photovoltaic grid-connected inverter

Abstract

The invention relates to the technical field of control methods and systems of photovoltaic grid-connected inverters, discloses a control method and system of a photovoltaic grid-connected inverter, and aims to solve the problems of maximum power point tracking, electric energy quality and system stability collaborative optimization of an inverter under a dynamic working condition in the prior art. According to the method and system, multi-dimensional data are collected in real time, and a space-time irradiance map is predicted based on image data; performing disturbance identification on power grid impedance; generating a global maximum power point voltage track in combination with irradiance prediction; and a multi-target model prediction control framework is constructed, and a system model containing power grid impedance is used for prediction. Through adoption of the technical scheme, prospective tracking of the global optimal power point is realized, the characteristics of the power grid are sensed in real time so as to adaptively guarantee the system stability and the electric energy quality, the power benefit, the electric energy quality, the system stability and the service life of devices are cooperatively optimized, and the comprehensive performance and the environmental adaptability of the system are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation technology, and specifically relates to a control method and system for photovoltaic grid-connected inverters. Background Technology

[0002] The field of renewable energy technology includes multiple branches such as photovoltaic power generation, wind power generation, and hydropower generation. The core of this technology is to convert natural energy into electrical energy to support the operation of modern power systems. Photovoltaic power generation refers to the technology of directly converting solar energy into direct current electricity using the photovoltaic effect of semiconductor materials. With the advancement of energy structure transformation, photovoltaic power generation technology has been widely applied, covering multiple application scenarios such as centralized power plants, distributed rooftops, and microgrids. The development of this field has not only promoted the innovation of energy technology, but also played an increasingly important role in optimizing global energy supply and ensuring the safety and stability of the power grid.

[0003] The control method for photovoltaic grid-connected inverters refers to the technology of precisely regulating the core power electronic equipment connecting the photovoltaic array to the public power grid. This method addresses the inherent intermittency and volatility of photovoltaic energy, as well as the complex and variable nature of the power grid environment, proposing an efficient and stable control strategy. By adjusting the on / off state of the power switching devices inside the inverter, this method achieves the conversion of DC power to high-quality AC power, ensuring that the output current is synchronized with the grid voltage in frequency and phase. In this method, the control system dynamically adjusts control commands by acquiring parameters from the grid and photovoltaic side in real time, thereby achieving core functions such as maximum power point tracking, power quality management, and grid support. Furthermore, this method improves the system's response speed and robustness under various operating conditions by optimizing the control algorithm.

[0004] Existing technologies for grid-connected inverter control mostly employ PI control strategies based on rotating coordinate systems. While their performance is acceptable under ideal grid conditions, they struggle to effectively suppress harmonic components of the output current when grid voltage distortion or imbalance occurs, leading to a decline in grid-connected power quality. The control system's response speed to dynamic disturbances such as sudden changes in sunlight intensity or grid faults is slow, easily triggering current overshoot and system oscillations, threatening the stability of the grid connection point. Traditional controller parameter tuning relies on precise system models; when grid impedance changes or system parameter perturbations occur, control performance deteriorates significantly, lacking robustness. Analysis of grid fault support capabilities is insufficient, especially during voltage dips, lacking proactive reactive power compensation and voltage support strategies, making it difficult to meet the low-voltage ride-through requirements of modern grids for renewable energy generation. Furthermore, incomplete decoupling control of active and reactive power leads to interference between them, limiting the flexibility and accuracy of power regulation. This problem is particularly pronounced in weak grid environments, directly impacting the overall grid-connected performance and grid friendliness of photovoltaic power generation systems. Summary of the Invention

[0005] The purpose of this invention is to provide a control method and system for a photovoltaic grid-connected inverter, aiming to solve the problem in the prior art of the difficulty in coordinating and optimizing the maximum power point tracking efficiency, power quality and system stability of photovoltaic grid-connected inverters under dynamic irradiance and grid operating conditions.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a control method for a photovoltaic grid-connected inverter, comprising the following steps:

[0007] S1: Real-time acquisition of voltage and current data of photovoltaic array, grid voltage and current data of grid connection point, temperature sensor data installed on inverter power devices, and sequence image data acquired by all-sky optical imaging sensors deployed in photovoltaic power plants, and integration to generate a multi-dimensional real-time dataset containing electrical, thermal and environmental conditions.

[0008] S2: Based on the sequence image data in the multidimensional real-time dataset, a pre-defined convolutional long short-term memory network model is used to extract spatiotemporal features and predict the dynamic evolution of cloud morphology, motion vectors and transparency, generating a spatiotemporal irradiance prediction map of the photovoltaic array surface covering a pre-defined future time window.

[0009] S3: Based on the grid voltage and current data of the grid connection point in the multidimensional real-time dataset, a preset, non-interference pseudo-random binary sequence disturbance signal is injected into the inverter output current reference command, and the grid connection point voltage response caused by the disturbance signal is collected synchronously. The voltage response and the frequency spectrum ratio of the disturbance signal are calculated by using the fast Fourier transform algorithm, thereby identifying the grid impedance spectrum data of the grid connection point within the preset frequency band.

[0010] S44: Call the spatiotemporal irradiance prediction map, combine it with the preset electrical characteristic model of the photovoltaic array, calculate the local optimal operating point on each spatial grid and future time step defined by the prediction map, and then solve the global maximum power point voltage trajectory to maximize the total output power at the entire array level through a global optimization algorithm.

[0011] S5: Construct a multi-objective model predictive control framework. This framework uses the inverter-grid coupled system state-space model containing the grid impedance spectrum data as the prediction model, and the reference power derived from the global maximum power point voltage trajectory as the core optimization objective. At the same time, it integrates multiple performance indicators such as grid current total harmonic distortion, DC bus voltage fluctuation amplitude, and power device thermal stress. Through online rolling optimization, it generates a set of optimal pulse width modulation drive signals to control the power switching transistors of the inverter.

[0012] As a further aspect of the present invention, the multidimensional real-time dataset includes: the DC voltage value of the photovoltaic array, the DC current value of the photovoltaic array, the effective and instantaneous values ​​of the three-phase AC voltage at the grid connection point, the effective and instantaneous values ​​of the three-phase AC current at the grid connection point, the housing temperature of the insulated gate bipolar transistor (IGBT) module, the housing temperature of the DC bus filter capacitor, and a timestamp-aligned digital image sequence continuously acquired at a frequency of 1 frame per second by a complementary metal-oxide-semiconductor (CMOS) image sensor with a center wavelength of 550 nanometers and a resolution of 1920x1080 pixels.

[0013] As a further aspect of the present invention, the spatiotemporal irradiance prediction map is a three-dimensional data matrix, with dimensions of spatial location x-coordinate, spatial location y-coordinate, and future prediction time t. Each element in the matrix represents the predicted composite value of direct solar radiation intensity (DNI) and diffuse solar radiation intensity (DHI) at a specific physical location of the photovoltaic array at a specific future time. The length of the future time pane is set to 60 seconds, with a time resolution of 1 second.

[0014] As a further aspect of the invention, the grid impedance spectrum data is a complex vector characterizing the grid impedance amplitude and phase angle characteristics from the grid base frequency (50 Hz) to one-tenth of the inverter switching frequency (e.g., 2 kHz). The amplitude of the pseudo-random binary sequence disturbance signal is set to 0.1% of the inverter's rated output current to ensure its imperceptibility to the grid.

[0015] As a further aspect of the present invention, the global maximum power point voltage trajectory is a time series vector, where each element corresponds to the optimal DC bus reference voltage at a time step within a future time pane. The generation of this trajectory avoids getting trapped in local maximum power points under local shading conditions, achieving forward-looking locking of the global optimal solution.

[0016] As a further aspect of the present invention, the steps for generating the spatiotemporal irradiance prediction map are as follows:

[0017] S211: Preprocess the sequence image data, including image geometric correction to eliminate fisheye lens distortion, and radiometric calibration to establish a mapping relationship between pixel grayscale values ​​and actual solar radiation intensity.

[0018] S212: Input the preprocessed image sequence into a pre-trained offline Convolutional Long Short-Term Memory (ConvLSTM) network. The convolutional layers of the network are used to extract the spatial distribution features of cloud textures, edges, and cloud blocks in each frame of the image; the long short-term memory layers of the network are used to learn and model the dynamic changes of these spatial features over time, i.e., the speed, direction, and shape evolution of cloud movement.

[0019] S213: The network outputs a prediction of a future image sequence. Through the established pixel grayscale-radiation intensity mapping relationship, the predicted future image sequence is converted into a quantized irradiance value and mapped to the physical coordinate system of the photovoltaic array, thus forming the spatiotemporal irradiance prediction map.

[0020] As a further aspect of the present invention, the identification step of the power grid impedance spectrum data specifically includes:

[0021] S311: The central controller of the inverter generates an m-sequence pseudo-random binary sequence (PRBS) as the reference signal i for the disturbance current. pert (t).

[0022] S312: Transfer the disturbance current reference signal i pert (t) and the conventional grid-connected current reference signal i grid (t) are added together to form the final current command signal, which is then used by the current controller to drive the inverter to inject a current i containing this tiny disturbance into the grid. grid (t).

[0023] S313: Utilizes a high-speed synchronous analog-to-digital converter to acquire the grid-connected point voltage v over a time period T. pcc (t) and grid-connected current i grid The waveform of (t).

[0024] S314: Process the acquired voltage and current waveforms separately, and extract the voltage response component Δv(t) and current response component Δi(t) caused by the disturbance through bandpass filtering.

[0025] S315: Perform Fast Fourier Transform (FFT) on Δv(t) and Δi(t) respectively to obtain their frequency domain representations V(f) and I(f).

[0026] S316: By calculating complex number division Z g (f) = V(f) / I(f), thus obtaining the grid impedance spectrum Z in the frequency band of interest. g (f) and stored in the inverter's memory for updating the system model within the multi-objective model predictive control framework.

[0027] As a further aspect of the present invention, the step of generating the global maximum power point voltage trajectory specifically includes:

[0028] S411: Establish a model describing the topology of a photovoltaic array, which includes the series and parallel connections of each photovoltaic module in the array and its physical coordinates.

[0029] S412: For each future time point t in the spatiotemporal irradiance prediction map k The spatial distribution data of irradiance at this moment I(x,y,t) k Load it into the array topology model.

[0030] S413: Based on the standard mathematical model of a single photovoltaic module (single diode model), combined with the different irradiance I(x,y,t) received by each module. k Using temperature data obtained from temperature sensors, the total output power of the entire photovoltaic array under any given DC bus voltage is calculated, thereby generating the PV characteristic curve of the array at that moment.

[0031] S414: Using particle swarm optimization or a genetic algorithm, perform a global search on the PV characteristic curve to determine the global maximum power point voltage V that maximizes the total output power. mpp (t k ).

[0032] S415: Repeat steps S412 to S414, iterating through all time points within the future time pane to form a V mpp (t k The sequence formed by ) is the global maximum power point voltage trajectory.

[0033] As a further aspect of the present invention, the execution steps of the multi-objective model prediction control framework are specifically as follows:

[0034] S511: At the beginning of each control cycle, update the discrete state-space model of the inverter-grid coupling system. The A and B matrix parameters of this model are based on the latest identified grid impedance spectrum data Z. g (f) Dynamically adjusted to accurately reflect the dynamic characteristics of the current power grid. The state vector x(k) includes the inductor current and capacitor voltage of the LCL filter.

[0035] S512: Define a multi-objective cost function J to be minimized. The cost function J is the sum of multiple weighted performance index terms. Its structure is as follows:

[0036] J = w1·∑(P) ref -P out ) 2 +w2·∑(THD i ) 2 +w3·∑(Δv dc ) 2 +w4·∑f(T j ).

[0037] S513: Explanation of each term in the cost function J: The first term is the power tracking error term, where P ref The reference power P at the current moment is calculated based on the global maximum power point voltage trajectory. out The first term is the predicted inverter output power; the second term is the grid-connected power quality term, where THD... i The first term is the predicted total harmonic distortion of the grid-connected current; the second term is the DC-side stability term, where Δv dc It is the difference between the predicted DC bus voltage and its rated value; the fourth item is the power device health item, where f(T) j () is a power device junction temperature T based on prediction. j It is a nonlinear function of its rate of change, used to quantify thermal stress.

[0038] S514: The weighting coefficients w1, w2, w3, and w4 are dynamically adjusted by an upper-level supervisory controller based on the macroscopic state of the system operation. When the spatiotemporal irradiance prediction spectrum indicates severe power fluctuations, w3 is increased to prioritize DC-side stability; when the grid impedance spectrum data Z g (f) When the power grid is highly inductive or capacitive in a certain harmonic frequency band, increase w2 to actively suppress harmonic injection at the corresponding frequency.

[0039] S515: Under the condition of satisfying the safety operation constraints of inverter current, voltage and power device temperature, solve the optimal control sequence u*(k), u*(k+1),...,u*(k+N-1) by numerical optimization algorithm (such as quadratic programming) to minimize the sum of cost function J in the next N control cycles.

[0040] S516: Only the first element u*(k) of the optimal control sequence (i.e., the optimal inverter arm output voltage at the current moment) is converted into a specific space vector pulse width modulation (SVPWM) signal and applied to the power switch of the inverter. In the next control cycle, the entire process from S511 to S516 is repeated, forming a rolling time-domain optimized closed-loop control.

[0041] A control system for a photovoltaic grid-connected inverter, used to execute the above method, the system comprising:

[0042] The multidimensional data acquisition module is equipped with interfaces that connect to the photovoltaic array, grid connection point, power devices, and all-sky optical imaging sensor, respectively, for synchronously acquiring photovoltaic array voltage and current, grid voltage and current, device temperature, and sky sequence images, and generating the multidimensional real-time dataset.

[0043] The spatiotemporal irradiance prediction module contains a convolutional long short-term memory network model, which is used to process the sequential image data in the multidimensional real-time dataset. By learning and inferring the dynamic spatiotemporal characteristics of the cloud layer, it generates a spatiotemporal irradiance prediction map of the photovoltaic array surface.

[0044] The grid impedance identification module includes a pseudo-random binary sequence signal generator and a fast Fourier transform processing unit, which is used to actively inject small disturbances and analyze the grid response to calculate and output the grid impedance spectrum data of the grid connection point in real time.

[0045] The forward-looking power optimization module, coupled with the spatiotemporal irradiance prediction module, and with a built-in topology and electrical characteristic model of the photovoltaic array, is used to calculate and generate the global maximum power point voltage trajectory based on the predicted non-uniform irradiance distribution through a global optimization search algorithm.

[0046] The multi-objective model predictive control module has a model predictive controller at its core. This controller is based on a system state-space model that is dynamically updated according to the grid impedance spectrum data. Under the constraints of a weighted cost function that includes multiple objectives such as power point tracking, power quality, DC stability and device health, it performs rolling time-domain optimization calculations and finally generates and outputs a precise pulse width modulation signal to control the inverter power switching transistors.

[0047] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0048] This invention, by introducing an all-sky optical imaging sensor and a spatiotemporal prediction algorithm, transforms inverter control from a passive response to an active prediction mechanism. It enables the inverter to anticipate irradiance changes caused by factors such as cloud movement, allowing the Maximum Power Point Tracking (MPPT) algorithm to overcome the lag and limitations of traditional perturbation observation methods under local shadows, achieving forward-looking and continuous tracking of the globally optimal power point. Simultaneously, the online grid impedance identification technology proposed in this invention allows the inverter to perceive the dynamic characteristics of its connected grid in real time. This provides a crucial basis for the adaptive adjustment of the control algorithm, ensuring high stability and excellent power quality even in weak or complex grid environments. Most importantly, this invention constructs a multi-objective model predictive control framework that unifies multiple previously conflicting or independently controlled objectives, such as power efficiency, power quality, system stability, and device lifespan, into a collaborative optimization mathematical framework. By dynamically adjusting the weights of each objective, it achieves optimal performance trade-offs under different operating conditions. For example, it maximizes power generation benefits when the weather is fine and the grid is strong, while prioritizing the safe and stable operation of the system during sudden weather changes or grid disturbances. This significantly improves the overall performance, intelligence level, and environmental adaptability of the photovoltaic grid-connected system. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the overall process of a photovoltaic grid-connected inverter control method provided in an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of the functional modules of a photovoltaic grid-connected inverter control system provided in an embodiment of the present invention;

[0051] Figure 3 for Figure 1 A detailed flowchart illustrating the steps involved in generating a spatiotemporal irradiance prediction map.

[0052] Figure 4 for Figure 1 A detailed flowchart illustrating the steps for identifying power grid impedance spectrum data.

[0053] Figure 5 This is a schematic diagram of the interaction process of the multi-objective model prediction control framework provided in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the description of this invention, it should be understood that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0055] Reference Figure 1 The present invention provides a control method for a photovoltaic grid-connected inverter, comprising the following steps:

[0056] Step S1: Real-time acquisition of voltage and current data of photovoltaic array, grid voltage and current data of grid connection point, temperature sensor data installed on inverter power devices, and sequence image data acquired by all-sky optical imaging sensors deployed in photovoltaic power plants, and integration to generate a multi-dimensional real-time dataset containing electrical, thermal and environmental conditions.

[0057] Step S2: Based on the sequence image data in the multidimensional real-time dataset, a pre-defined convolutional long short-term memory network model is used to extract spatiotemporal features and predict the dynamic evolution of cloud morphology, motion vectors and transparency, generating a spatiotemporal irradiance prediction map of the photovoltaic array surface covering a pre-defined future time window.

[0058] Step S3: Based on the grid voltage and current data of the grid connection point in the multidimensional real-time dataset, a preset, non-interference pseudo-random binary sequence disturbance signal is injected into the inverter output current reference command, and the grid connection point voltage response caused by the disturbance signal is collected synchronously. The voltage response and the frequency spectrum ratio of the disturbance signal are calculated by using the fast Fourier transform algorithm, thereby identifying the grid impedance spectrum data of the grid connection point within the preset frequency band.

[0059] Step S4: Call the spatiotemporal irradiance prediction map, combine it with the preset electrical characteristic model of the photovoltaic array, calculate the local optimal operating point on each spatial grid and future time step defined by the prediction map, and then solve the global maximum power point voltage trajectory to maximize the total output power at the entire array level through a global optimization algorithm.

[0060] Step S5: Construct a multi-objective model predictive control framework. This framework uses the inverter-grid coupled system state-space model containing the grid impedance spectrum data as the prediction model, and the reference power derived from the global maximum power point voltage trajectory as the core optimization objective. At the same time, it integrates multiple performance indicators such as grid current total harmonic distortion, DC bus voltage fluctuation amplitude, and power device thermal stress. Through online rolling optimization, a set of optimal pulse width modulation drive signals is generated to control the inverter's power switching transistors.

[0061] In one specific embodiment, the process of generating the multidimensional real-time dataset in step S1 is described in detail. This process is broken down into multiple parallel sub-steps, coordinated by a central data acquisition and synchronization unit.

[0062] Step S101: Synchronously acquire electrical status data. Specifically, deploy a high-precision DC voltage sensor and a shunt or Hall effect current sensor at the output of the DC combiner box of the photovoltaic array to measure the total output DC voltage V of the photovoltaic array. pv With DC current I pv Between the inverter's AC output side and the grid's common coupling point, high-bandwidth voltage transformers and current transformers are configured for each phase of the three-phase system to measure the instantaneous value v of the three-phase grid-connected voltage. pcc_a (t),v pcc_b (t),v pcc_c (t) and the instantaneous value of the three-phase grid-connected current i grid_a (t),igrid_b (t),i grid_c (t). All electrical quantity sensors are sampled via a multi-channel synchronous analog-to-digital converter (ADC).

[0063] In one example, the ADC's sampling frequency is set to 20kHz with a 16-bit resolution. This sampling frequency is set based on the principle that it must be significantly higher than the grid base frequency (50Hz or 60Hz) and at least twice the inverter switching frequency (e.g., 10kHz) to satisfy the Nyquist theorem and ensure distortion-free reconstruction of high-frequency harmonics and disturbance signals. The 16-bit resolution guarantees measurement accuracy. For example, at a certain sampling time k, the acquired raw data might be: V pv =652.18V, I pv =125.4A, v pcc_a (k) = 320.5V, i grid_a (k) = 85.2A. These instantaneous values ​​are stored in a first-in-first-out (FIFO) buffer, awaiting synchronization and encapsulation with other types of data.

[0064] Step S102 involves acquiring thermal state data. Specifically, negative temperature coefficient (NTC) thermistors are tightly mounted on the heat sinks of key power devices inside the inverter, such as the IGBT modules, and on the surface of the DC bus support capacitor housing. The resistance of these thermistors changes with temperature. A dedicated low-speed data acquisition unit samples the voltage signals of these thermistors at a frequency of 10Hz.

[0065] The sampling frequency is set based on the inertia of thermodynamic systems. The rate of temperature change in power devices and capacitors is much lower than the rate of change in electrical quantities; a sampling frequency of 10Hz is sufficient to capture their dynamic temperature rise process. The acquired voltage signal is converted using a preset lookup table or by applying the Steinhart-Hart equation to obtain an accurate temperature value. For example, the case temperature of IGBT module A at a certain moment might be T. igbt,A =82.5℃, the average temperature of the DC bus capacitor bank is T cap =61.3℃.

[0066] As an anomaly handling mechanism, the acquisition unit continuously monitors whether the output of the thermistor exceeds the normal range (e.g., resistance close to 0 or infinity), which typically indicates a short circuit or open circuit fault in the sensor. Once a fault is detected, the system generates an alarm and marks the corresponding data point as invalid. Simultaneously, the control system can selectively activate a junction temperature estimation algorithm based on a device current and switching loss model as a backup data source to ensure that the system's thermal protection function is not interrupted.

[0067] Step S103 involves acquiring a sequence of environmental images, specifically a sky image sequence. Specifically, an all-sky optical imaging sensor is installed at the center or highest point of the photovoltaic power station. This sensor is an industrial-grade camera integrating a fisheye lens, with a complementary metal-oxide-semiconductor (CMOS) image sensor at its core.

[0068] In one specific embodiment, the CMOS sensor has a physical resolution of 1920x1080 pixels and its spectral response is optimized with a center wavelength at 550 nanometers, a band highly sensitive to the optical characteristics of clouds. The camera continuously captures sky images at a fixed frequency of one frame per second. This frequency setting is a trade-off: too low a frequency fails to capture details of rapidly moving clouds, while too high a frequency results in redundant data and increases the computational burden of subsequent processing. One frame per second has proven sufficient for minute-level irradiance prediction tasks. The captured raw image data (e.g., 12-bit RAW format) is transmitted in real-time to an edge computing unit.

[0069] Step S104 involves the integration and synchronization of multidimensional data. All acquisition units (electrical, thermal, and environmental) are connected to a common time synchronization network. This network is based on Precise Time Protocol (PTP) or Network Time Protocol (NTP), with a unified time reference provided by a GPS time server, ensuring that the timestamp error of all data is within the microsecond level. At each macroscopic time step of the control system (e.g., 1 second), the central data acquisition and synchronization unit performs the following operations: extracts all sampled data from the electrical quantity FIFO buffer within the past second; obtains the latest temperature reading from the thermal data acquisition unit; and obtains the latest frame of sky image from the edge computing unit. Subsequently, this heterogeneous data is encapsulated into a unified multidimensional real-time dataset structure.

[0070] The dataset structure can be defined as a record containing the following fields:

[0071] {

[0072] master_timestamp:uint64_t, / / UTC timestamp with nanosecond precision

[0073] pv_voltage:float32, / / Average DC voltage of the photovoltaic array

[0074] pv_current:float32, / / Average DC current of the photovoltaic array

[0075] grid_voltage_waveform: array[float32], / / Sequence of instantaneous three-phase voltage values ​​at the grid connection point

[0076] grid_current_waveform:array[float32], / / Instantaneous sequence of three-phase current values ​​at the grid connection point

[0077] igbt_temperatures:map<string,float32> / / Temperature of each IGBT module

[0078] cap_temperature:float32, / / DC bus capacitor temperature

[0079] sky_image_path: string / / Path to the original sky image file stored at this moment

[0080] }

[0081] This dataset provides a unified, synchronized, and comprehensive input for all subsequent prediction and control algorithms.

[0082] Reference Figure 3 The process of generating the spatiotemporal irradiance prediction map in step S2 is described in detail. This process uses the sequence image data obtained in step S1 to predict the irradiance changes in the near future through a deep learning model.

[0083] Step S201 involves preprocessing the acquired sequence image data. This step is crucial for ensuring the quality and consistency of the input data for subsequent prediction models.

[0084] Step S2011: Perform image geometric correction. Due to the use of a fisheye lens, the original image suffers from severe barrel distortion, causing linear features in the sky (such as aircraft contrails) to appear as curves. To eliminate this distortion, the system performs a one-time offline calibration beforehand. During calibration, a large checkerboard calibration board is used to capture images at different positions and angles within the camera's field of view. By analyzing the known physical coordinates of the checkerboard corner points and their pixel coordinates in the image, the camera's intrinsic parameter matrix K and distortion coefficient vector D = [k1,k2,p1,p2,k3] are calculated using the Zhang Zhengyou calibration method. These parameters are stored in the system. During real-time operation, each newly acquired image frame is input into a distortion correction function. This function uses the stored K and D matrices to remap each pixel of the original image to its position under a distortion-free ideal pinhole camera model, generating a geometrically accurate image.

[0085] Step S2012: Perform image radiometric calibration. This step aims to establish the grayscale value of the pixel relative to the actual solar radiation intensity received by the ground (unit: W / m²). 2A quantitative mapping relationship between the two is established. The calibration process requires deploying one or more high-precision reference pyranometers within the photovoltaic power station. During the calibration period, which lasts for weeks or even months, the system synchronously records sky images and pyranometer readings. By analyzing a large number of data pairs under various weather conditions (sunny, cloudy, overcast), a mapping model is built. Where p is the pixel gray value corresponding to the sky region in the image, θ z and These are the zenith angle and azimuth angle of the sun, respectively. The function f can be a multinomial regression model or a small feedforward neural network. Its parameters are obtained by least-squares fitting to a calibration dataset and stored in the system. This mapping relationship allows changes in cloud brightness in an image to be directly converted into changes in surface irradiance.

[0086] Step S202: The preprocessed image sequence is input into a pre-trained offline Convolutional Long Short-Term Memory (ConvLSTM) network for spatiotemporal evolution prediction. The structure and operation of this network are as follows:

[0087] The network input is a time-series tensor with dimensions [N, T]. in [,H,W,C], where N is the batch size, T in H is the length of the input sequence (e.g., 10 frames of images from the past 10 seconds), H and W are the height and width of the preprocessed image (e.g., downsampled to 256x256), and C is the number of channels (e.g., 1 grayscale channel).

[0088] The network consists of multiple stacked ConvLSTM layers. In a specific implementation, the network comprises three ConvLSTM layers. The first layer uses 64 7x7 convolutional kernels to capture large-scale cloud features; the second layer uses 128 5x5 convolutional kernels to learn medium-scale cloud textures and edges; and the third layer uses 64 3x3 convolutional kernels to capture fine-grained dynamic details of the clouds. The convolutional operations within the ConvLSTM units enable them to directly process spatial data, while their recurrent structure (including input gates, forget gates, and output gates) allows them to learn and remember the temporal variations of these spatial features, namely cloud movement, formation, and deformation.

[0089] The network's output layer is a 1x1 convolutional layer that remaps the high-dimensional feature map output from the last ConvLSTM layer back to a linear convolutional layer with dimensions [N,T]. out Tensors of [H,W,C], where T out It is the predicted future time step (e.g., predicting 60 frames of images in the next 60 seconds).

[0090] The ConvLSTM model was trained offline. The training dataset consisted of a series of historical sky images spanning several years, along with their corresponding future true irradiance maps measured by an in-field radiometer network. The Adam optimizer and mean squared error (MSE) loss function were used to optimize the network weights through backpropagation until the model's prediction error on the validation set converged.

[0091] Step S203: Convert the output of the ConvLSTM network into the final spatiotemporal irradiance prediction map of the photovoltaic array surface.

[0092] Step S2031: Convert the predicted future image sequence into a quantized irradiance map. For each frame of the predicted image output by the network, apply the inverse function of the radiometric calibration function f established in step S2012. The predicted grayscale value of each pixel is converted into the corresponding predicted solar radiation intensity value (W / m²). 2 ).

[0093] Step S2032 involves projecting the sky's irradiance map onto the physical plane of the photovoltaic array. This process requires a pre-calculated geometric projection model. This model is based on the camera's precise installation location (latitude and longitude, altitude), orientation (azimuth and elevation angles), and the precise three-dimensional geographic coordinates of each photovoltaic module in the array. For any future time t_k, the system first calculates the sun's position (zenith angle and azimuth angle) at that time. Then, for each pixel in the sky image, the direction of the sunlight it represents can be calculated. Through ray tracing, the intersection point of this ray with the photovoltaic array plane is calculated, thereby establishing a mapping relationship between the sky image pixel coordinates (u,v) and the array's physical coordinates (x,y). The irradiance values ​​obtained in step S2031 are then filled into a two-dimensional grid representing the photovoltaic array according to this mapping relationship, forming the irradiance spatial distribution map for that time t_k.

[0094] Step S2033: Integrate the predictions from all future time steps to form the final three-dimensional data matrix. Repeat step S2032, iterating through all time points within the future time pane (e.g., from t+1 seconds to t+60 seconds), generating 60 two-dimensional irradiance spatial distribution maps. Stack these maps along the time axis to finally form a three-dimensional spatiotemporal irradiance prediction map I(x,y,t). The dimensions of this map are [GridSize_X,GridSize_Y,T]. out ], where GridSize_X and GridSize_Y are the number of grid divisions in the array plane, and T out This is the predicted duration. Each element I(i,j,k) in the matrix represents the predicted solar radiation intensity at the (i,j)th grid position in the array at the kth second in the future.

[0095] As an alternative, step S2 can also be implemented using a deterministic method based on classical computer vision techniques. This alternative includes:

[0096] Step A1: Calculate cloud motion vectors using optical flow. For two consecutive preprocessed frames of sky images, apply a dense optical flow algorithm (e.g., Gunnar Farneback algorithm). This algorithm calculates a two-dimensional motion vector (vx, vy) for each pixel in the image, thus forming a complete cloud motion vector field.

[0097] Step A2 involves smoothing and segmenting the motion vector field. Noise and outlier vectors are removed using methods such as median filtering. Then, clustering algorithms (such as K-means) can be used to group pixels with similar motion vectors, identifying different cloud clusters with consistent motion trends.

[0098] Step A3 involves trajectory extrapolation and shadow prediction. For each cloud cluster, its trajectory within a future time pane is extrapolated linearly or non-linearly based on its current position and average motion vector. Subsequently, using the geometric projection model from step S2032, these predicted future cloud cluster positions are projected onto the photovoltaic array plane to generate a shadow distribution map for future times.

[0099] Step A4: Generate an irradiance map using a clear-sky model. Calculate the irradiance under cloudless conditions using a clear-sky radiation model. Within the predicted shadowed area, attenuate the clear-sky irradiance based on the cloud's optical thickness (estimated from pixel grayscale values) to obtain the final irradiance prediction. This alternative approach has lower computational complexity than ConvLSTM, but its prediction accuracy is highly dependent on the smoothness and predictability of cloud movement, and it is less capable of modeling rapid changes in cloud shape or cloud formation / disappearance processes.

[0100] As an alternative, step S2 can be based on satellite cloud imagery data for prediction. This approach is suitable for scenarios requiring a longer prediction timeframe (e.g., 15 minutes to several hours).

[0101] Step B1: Obtain cloud imagery data from geostationary meteorological satellites. This involves periodically acquiring satellite cloud images covering the area where the photovoltaic power station is located from a meteorological service provider via a web API, for example, acquiring infrared or visible light channel images every 15 minutes.

[0102] Step B2 involves applying the CloudMotionVector algorithm or a similar machine learning model (such as U-Net) to process the continuous sequence of satellite images to predict the evolution of the cloud imagery over the next few hours.

[0103] Step B3 converts the predicted satellite cloud image into surface solar radiation using a radiative transfer model. This model requires atmospheric parameters (such as aerosol and water vapor content) to calculate the attenuation of solar radiation by clouds. This approach has significantly lower spatial resolution (typically in the kilometer range) and temporal resolution (in the 15-minute range) than ground-based imaging methods, making it unsuitable for second-level inverter dynamic control. However, it can provide medium- to long-term forecast information for power plant-level power dispatch and energy storage management.

[0104] Reference Figure 4 The process of identifying grid impedance spectrum data in step S3 is described in detail. This process detects the dynamic characteristics of the grid by actively injecting small disturbances into the grid.

[0105] Step S301: A pseudo-random binary sequence (PRBS) signal is generated as a disturbance source. Specifically, this signal is generated by an m-sequence generator inside the inverter's digital signal processor (DSP). The m-sequence is a periodic signal with good autocorrelation properties (close to white noise) and determinism.

[0106] In one example, we choose a 10th-order m-sequence with a period length of 2. 10 -1 = 1023. The clock frequency f_clk of the sequence is set to 4kHz. This setting determines the spectral range and resolution of the perturbation signal. The spectral energy will be distributed in f_clk / (2 10 The disturbance frequency range extends from the harmonic frequency point of approximately 3.9 Hz (-1) to the Nyquist frequency f_clk / 2 = 2 kHz. This frequency range covers the grid fundamental frequency (50 Hz) to most of the harmonic frequencies of concern (such as the 3rd, 5th, 7th, 11th, and 13th harmonics) and the subsynchronous / supersynchronous oscillation bands that may interact with the inverter control loop. The amplitude of the PRBS signal is set to 0.1% of the inverter's rated output current. For example, for an inverter with a rated output current of 100 A, the peak-to-peak value of the disturbance current is only 0.2 A. Such a small amplitude ensures that the voltage fluctuations caused by the injected disturbance at the grid point of common coupling (PCC) are minimal and imperceptible to the grid and other paralleled devices, meeting the power quality requirements in the grid connection guidelines.

[0107] Step S302: The generated disturbance signal is injected into the power grid. The PRBS signal i pert (t) Current reference command i output by the inverter's conventional dq-axis current controller in the digital domain. grid (t) are added together to form the final current command i. final (t)=i grid (t)+i pert(t). The inverter's inner-loop current controller will precisely track this final command, controlling the PWM duty cycle of the power switch transistors to ensure that the actual current i output by the inverter to the grid is... grid (t) contains this small, wideband PRBS disturbance component. This injection method is typically performed on the d-axis current reference, meaning the disturbance mainly manifests as a small fluctuation in the active current.

[0108] Step S303: Perform synchronous data acquisition of the disturbance response. Specifically, using the same multi-channel synchronous analog-to-digital converter as in step S101, during the disturbance injection period, sample the instantaneous three-phase voltage v at the grid connection point at the same high sampling frequency (e.g., 20kHz). pcc (t) and instantaneous value of three-phase current i grid (t) Continuous sampling is performed. The acquisition time T is set to one or more complete periods of the m-sequence to ensure stable spectral estimation in subsequent frequency domain analysis. For example, for an m-sequence with a period of 1023 and a clock frequency of 4kHz, the period duration is approximately 1023 / 4000Hz ≈ 0.256 seconds. A typical acquisition window T can be set to 4 periods, or approximately 1.024 seconds, to acquire enough data points for averaging and improve the signal-to-noise ratio. The acquired raw voltage and current waveform data are stored in a dedicated memory buffer for subsequent processing.

[0109] Step S304 involves processing the acquired waveform data to extract the response components. The purpose of this step is to accurately separate the voltage and current response components caused by weak PRBS perturbations from a strong signal containing the fundamental frequency and background harmonics. One specific implementation method utilizes signal subtraction techniques. Due to the injected PRBS perturbation sequence i pert (t) is known, and the system can obtain the total current i from the collected data. grid Subtracting the fundamental frequency and main harmonic components estimated by the phase-locked loop (PLL) from (t) yields the current response component Δi(t). Furthermore, a cross-correlation algorithm can be used to calculate the acquired grid-connected point voltage v. pcc The cross-correlation function between (t) and the original PRBS sequence used as a reference signal, and the acquired grid-connected current i grid(t) Cross-correlation function with the PRBS sequence. Since the PRBS signal has white noise-like autocorrelation characteristics (approaching zero outside of zero delay) and is uncorrelated with the fundamental frequency and harmonic components of the power grid itself, cross-correlation can effectively suppress noise and uncorrelated signals, extracting the system impulse response synchronized with the disturbance. The system frequency response can be obtained by performing a Fourier transform on the impulse response. As a more direct alternative, a digital bandpass filter can be used. Design a set of narrowband bandpass filters with center frequencies aligned with the harmonic frequencies in the PRBS disturbance signal spectrum. The acquired v... pcc (t) and i grid (t) By passing through this set of filters, the voltage response components Δv(f) at each disturbance frequency point can be extracted. n ) and current response component Δi(f n ).

[0110] Step S305: Perform a Fast Fourier Transform (FFT) on the extracted response components. The voltage response time-domain sequence Δv(t) and the current response time-domain sequence Δi(t), both of length T obtained after step S304, are input into the FFT calculation unit. Before performing the FFT, a window function, such as a Hanning window or a flat-top window, is typically applied to the time-domain sequences. The purpose of applying the window function is to reduce spectral leakage caused by signal truncation and improve frequency resolution and the accuracy of amplitude estimation. The length N of the FFT is usually chosen to be an integer power of 2 and greater than or equal to the number of sampling points to utilize the algorithm's efficiency. The result of the FFT operation is the complex spectrum V(f) of the voltage response and the complex spectrum I(f) of the current response. Each element of these two spectral vectors corresponds to a specific frequency point and contains the amplitude and phase information of the signal at that frequency point.

[0111] Step S306: Calculate and store the grid impedance spectrum. Based on Ohm's law in the frequency domain, the grid impedance Z can be obtained by performing complex division on the voltage spectrum V(f) and current spectrum I(f) obtained in step S305. g (f). The specific calculation formula is Z. g (f) = V(f) / I(f). This calculation is performed at all frequencies of interest, ultimately generating a complex vector, namely the grid impedance spectrum data. Each element Z of this vector... g (f n ) represents the power grid at frequency f n The impedance at point Z includes the amplitude |Z g (f n )| and phase angle ∠Z g (f n For example, the identification results might show that at 350Hz (7th harmonic), the power grid exhibits high inductive impedance, with a value of Z.g (350Hz) = 0.5 + j1.2 ohms. This impedance spectrum data is stored in the inverter controller's non-volatile memory and is assigned a valid timestamp. The impedance identification process is executed automatically periodically (e.g., every 5 minutes) or triggered after a significant disturbance in the grid voltage is detected, ensuring that the grid model used in the multi-objective model predictive control framework always reflects the latest dynamics of the grid.

[0112] Reference Figure 1 The process of generating the global maximum power point voltage trajectory in step S4 is described in detail, along with the detailed flowchart not shown in the attached figures.

[0113] Step S401: Establish a refined photovoltaic array topology and electrical characteristic model. This model is established during the system commissioning phase or through an automated scanning program and stored in the inverter's memory. Specifically, the model is a data structure containing detailed information about each photovoltaic module in the array. For each module, the model records its unique identifier, its physical three-dimensional coordinates (x, y, z) in the array, its substring number and position within the string, and the electrical characteristics of its bypass diodes. Furthermore, the model also includes key electrical characteristic parameters of this photovoltaic module model under standard test conditions (STC), such as the open-circuit voltage V. oc Short-circuit current I sc Maximum power point voltage V mp Maximum power point current I mp The model also includes the temperature coefficients of current and voltage. The granularity of the model can be refined to the individual photovoltaic cell level if more nuanced internal mismatch effects need to be considered.

[0114] Step S402: Load the predicted irradiance data into the array model. For each future time step t in the spatiotemporal irradiance prediction map I(x,y,t). k (k ranges from 1 to 60), the system performs one mapping operation. This generates a two-dimensional irradiance spatial distribution map I(x,y,t) at that moment. k Load the data. For each photovoltaic module in the model of step S401, based on its physical coordinates (x, y), extract the average solar radiation intensity value received on its surface from the irradiance distribution map I(x, y, t_k) using bilinear interpolation or nearest neighbor interpolation. Simultaneously, using the temperature sensor data collected in step S102, combined with a simplified thermal model, estimate the operating temperature of each module at that moment. Therefore, at time t... k Each component in the array is assigned a specific, possibly different, irradiance and temperature value.

[0115] Step S403: Generate the PV characteristic curve of the entire photovoltaic array at a specific time. Based on the standard five-parameter or seven-parameter (dual-diode) mathematical model of a single photovoltaic module, for each module in the array, using the specific irradiance and temperature values ​​assigned to it at time t_k, calculate its IV characteristic curve under that operating condition. Subsequently, according to the series-parallel topology defined in the model in step S401, the output characteristics of the entire photovoltaic array are synthesized numerically. Specifically, for a given DC bus voltage V... dc The system needs to solve a large system of nonlinear equations to determine the current in each substring and the total output current I. total Due to the presence of bypass diodes, when the irradiance between components is uneven (i.e., local shading occurs), the overall PV characteristic curve (P) of the array will be affected. total =V dc *I total This will result in multiple Local Maximum Power Points (LMPPs) and one Global Maximum Power Point (GMPP). This is achieved by scanning V within a preset voltage range (e.g., from 0 to the total open-circuit voltage of the array). dc By repeating the above calculation, the time t can be generated. k Below, the complete multi-peak PV characteristic curve of the photovoltaic array is shown.

[0116] Step S404 involves global optimization of the generated multi-peak PV characteristic curve. To accurately find the global maximum power point from multiple local extrema, a heuristic global optimization algorithm is employed. In a specific embodiment, the Particle Swarm Optimization (PSO) algorithm is used. The algorithm initializes a population of M particles, where the position of each particle represents a candidate DC bus voltage V_dc. The fitness function value of each particle is defined as the total array output power P calculated in step S403 at that voltage. total (V dc During the iteration process, each particle updates its velocity and position based on its own historical best position (pbest) and the global best position (gbest) of the entire population. After several iterations, all particles will converge to the global maximum power point on the PV curve. The voltage value V at this point is... mpp (t k That is, at the future time t k The optimal operating voltage can be determined. Alternatively, other global search algorithms such as Genetic Algorithm (GA), Differential Evolution (DE), or Simulated Annealing (SA) can be used.

[0117] Step S405: Iterate through the future time panes to generate voltage trajectories. Repeat steps S402 to S404 for each future time point t covered by the spatiotemporal irradiance prediction map. kFor each k (from 1 to 60), a complete array PV characteristic generation and global optimization calculation are performed. This will result in a sequence of 60 global maximum power point voltage values: [V mpp (t k+1 ),V mpp (t k+2 ),...,V mpp (t k+60 This time series vector is the final generated global maximum power point voltage trajectory. This trajectory provides a forward-looking, dynamic DC voltage reference target for the subsequent multi-objective model predictive control module, enabling it to adjust the inverter's operating point in advance to smoothly respond to upcoming irradiance changes, rather than passively waiting for power fluctuations to occur before searching for a new maximum power point.

[0118] Reference Figure 5 The execution process of the multi-objective model predictive control framework in step S5 is described in detail. This framework performs rolling optimization once within each control cycle (e.g., 50 microseconds).

[0119] Step S501: Update the prediction model of the inverter-grid coupled system. The core of the controller is a discrete-time state-space model used to predict the dynamic behavior of the system over the next N control cycles (i.e., the prediction time domain). This model is typically expressed as: x(k+1) = A d ·x(k)+B d ·u(k)+E d ·d(k). Here, the state vector x(k) contains variables characterizing the state of the system's energy storage elements. In a three-phase inverter using an LCL filter, x(k) typically includes the three-phase inductor current on the inverter side, the three-phase inductor current on the grid side, and the three-phase voltage of the filter capacitor. The control input vector u(k) is the three-phase voltage output from the inverter arm, a quantity directly controllable by the controller. The disturbance input vector d(k) represents uncontrollable external inputs, primarily referring to the grid voltage at the grid connection point. The key to this step is that the model matrices A_d and B_d are not fixed but dynamically adaptive. After each impedance identification (step S3), the system utilizes the latest grid impedance spectrum data Z. g (f) Updates the equivalent circuit parameters of the power grid (e.g., equivalent inductance and resistance) in the continuous-time domain model. Then, the updated continuous-time model is transformed into a new discrete state-space matrix A using an accurate discretization method (e.g., the zero-order hold method). d and B d This dynamic update mechanism enables the prediction model to accurately reflect the actual characteristics of the current power grid, such as impedance increases or resonance peaks under weak grid conditions.

[0120] Step S502: Construct and quantify the multi-objective cost function J. At the beginning of each control cycle, the controller constructs a cost function to minimize within the prediction time domain. This cost function is a weighted sum of multiple performance metrics, and its general form is:

[0121] J = ∑[w p ·(P ref (k+j)-P out (k+j)) 2 +w q ·Q err (k+j) 2 +w thd ·THD i (k+j) 2 +w v ·(Δv dc (k+j)) 2 +w t ·f(T j (k+j))]

[0122] The summation symbol Σ represents the accumulation of all prediction steps j from the current time k to the future k+N-1.

[0123] Step S503 provides a detailed explanation of each component of the cost function.

[0124] The first term is the power tracking term. P ref (k+j) is the reference active power at the j-th future step. This reference value is generated by the global maximum power point voltage trajectory V generated by the forward power optimization module (step S4). mpp The value was calculated based on the PV characteristics of the photovoltaic array. out (k+j) is the inverter output active power predicted using the state-space model in step S501. This is intended to drive the inverter to accurately track the maximum available solar energy.

[0125] The second item is the reactive power tracking item. Q err (k+j) is the error between the predicted reactive power output and the reference value. The reference value can be set by grid dispatch instructions or determined by the inverter's local voltage support function.

[0126] The third item is power quality. THD i (k+j) represents the predicted total harmonic distortion (THD) of the grid-connected current. This value is calculated online by performing a Fast Fourier Transform on the future current waveform predicted by the state-space model. This term is used to actively suppress harmonic injection and ensure grid-connected power quality.

[0127] The fourth term is the DC-side stability term. Δv dc (k+j) is the predicted DC bus voltage vdc Its reference value (i.e., V) mpp The deviation between (k+j)). The dynamics of the DC bus voltage are determined by an additional energy balance equation.

[0128]

[0129] Coupled with the main state-space model. This term is used to suppress DC bus voltage overshoot or sag caused by power surges, maintaining system stability.

[0130] The fifth item is the health status of power devices. f(T) j (k+j) is a function that quantifies thermal stress. The controller integrates a simplified power device junction temperature T. j Real-time predictive models (e.g., RC thermal network models) are used. The input to this model is the switching and conduction losses calculated based on the predicted current and switching states. The function f can be designed to penalize high junction temperature peaks and large junction temperature fluctuations (ΔT). j Both of these are key factors that lead to aging and failure of power modules.

[0131] Step S504: Perform dynamic adaptive adjustment of the weight coefficients. The weight coefficients w in the cost function... p ,w q ,w thd ,w v , w t It is not fixed, but dynamically adjusted by a higher-level monitoring logic based on the macroscopic state of the system to achieve the optimal performance trade-off under different operating conditions.

[0132] Specifically, when the spatiotemporal irradiance prediction map generated in step S2 indicates that a large cloud cover will block or move away within the next few tens of seconds, causing drastic changes in photovoltaic power, the supervisory logic will significantly increase the weight w of the all-DC side stability term. v and the weight w of the thermal stress term t At the same time, appropriately reduce the weight w of the power tracking term. p This causes the controller to prioritize measures (such as actively limiting the rate of power change) to ensure the stability of the DC bus voltage and reduce the thermal shock to the IGBT, even if this temporarily sacrifices some maximum power point tracking accuracy.

[0133] Conversely, under clear, cloudless conditions with stable irradiance, the supervisory logic will increase the weight w of the success rate tracking item. p In order to maximize power generation and economic benefits.

[0134] Furthermore, when the grid impedance spectrum Z identified in step S3 g(f) When the power grid exhibits a parallel resonance peak near a certain harmonic frequency (e.g., the 5th or 7th harmonic), the supervisory logic significantly increases the weight of the power quality term. thd Furthermore, a term specifically penalizing harmonic currents at that particular frequency can be added to the cost function. This allows the controller to proactively adjust its switching strategy to avoid injecting energy at that resonant frequency, thereby preventing harmonic amplification and system instability.

[0135] Step S505: Solve the constrained optimization problem online. The cost function J is expressed as a quadratic function of the future control sequence U = [u(k), u(k+1), ..., u(k+N-1)], and combined with the system's linear constraints, a standard constrained quadratic programming (QP) problem is formed. The constraints include: the inverter output current cannot exceed its rated value; the inverter output voltage cannot exceed the DC bus voltage limit; and the predicted power device junction temperature cannot exceed its safe operating upper limit. In each control cycle, an efficient embedded QP solver (e.g., based on the interior-point method or active-set method) is invoked to compute the optimal control sequence U* that minimizes the cost function J.

[0136] Step S506: Implement rolling time-domain control. Although step S505 calculates the optimal control sequence U* for the next N steps, according to the rolling time-domain principle of model predictive control, only the first element u*(k) of this sequence is actually used. This optimal voltage vector u*(k) is fed into the space vector pulse width modulation (SVPWM) module to generate precise gate switching signals for driving the six power switches in the three-phase bridge arm of the inverter. In the next control cycle (at time k+1), the system acquires new measurements, updates the state estimate, and repeats the entire process from S501 to S506, recalculating the optimization based on the new system state. This continuous closed-loop cycle of "prediction-optimization-implementation-remeasurement" enables the control system to continuously correct prediction errors and has strong robustness to unmodeled dynamics and disturbances.

[0137] Reference Figure 2 The present invention also provides a control system for a photovoltaic grid-connected inverter, which is designed to execute the above-described control method. The system includes:

[0138] A multi-dimensional data acquisition module integrates multiple physical interfaces. It connects to the DC side of the photovoltaic array via high-precision shunt and voltage divider interfaces; to the grid's common coupling point via voltage and current transformer interfaces; to the power devices and capacitors inside the inverter via a thermistor signal conditioning circuit interface; and to an all-sky optical imaging sensor deployed at the site via an Ethernet or USB interface. Internally, the module includes a synchronous clock unit with PTP protocol support and a high-speed multi-channel ADC, responsible for synchronously sampling, quantizing, and timestamping data from all sources, ultimately generating and outputting the multi-dimensional real-time dataset.

[0139] A spatiotemporal irradiance prediction module, which can be a dedicated hardware acceleration unit, such as an edge computing device embedded with a graphics processing unit (GPU) or a field-programmable gate array (FPGA), internally embeds an offline-trained Convolutional Long Short-Term Memory (ConvLSTM) network model. This module receives a sequence of sky images provided by a multidimensional data acquisition module, performs image preprocessing, spatiotemporal feature extraction, and dynamic evolution prediction, and ultimately generates and outputs a spatiotemporal irradiance prediction map of the photovoltaic array surface covering short-term future time panes.

[0140] A grid impedance identification module, typically implemented as a software function on the inverter's central digital signal processor (DSP), includes an m-sequence pseudo-random binary sequence (PRBS) signal generator logic and a dedicated hardware coprocessor or optimized software library for performing fast Fourier transforms. Based on a preset period or external triggering, this module controls the injection of disturbance signals, synchronously analyzes the voltage and current responses, and calculates and updates the grid impedance spectrum data stored in shared memory in real time.

[0141] A forward-looking power optimization module, also a software task running on the central controller, incorporates a database of refined topology and electrical characteristic models of the photovoltaic array. This module receives the predicted spectrum output from the spatiotemporal irradiance prediction module and, for each future time step, calculates the global maximum power point under the predicted non-uniform irradiance conditions by executing a global optimization search algorithm (such as particle swarm optimization). Finally, it generates and outputs the voltage trajectory of the global maximum power point.

[0142] A multi-objective model predictive control module, the core of the entire control system, is implemented on a high-performance DSP. Internally, this controller maintains a dynamically updated state-space model of the inverter-grid coupled system based on the output of the grid impedance identification module. In each control cycle, it obtains the latest reference commands (such as power trajectory) and system state from other modules, constructs a weighted cost function encompassing multiple objectives including power point tracking, power quality, DC stability, and device health, solves a constrained quadratic programming problem online, and ultimately generates and outputs a precise space vector pulse width modulation (SVPWM) signal, which controls the switching of the inverter's power switches via a gate drive circuit.

[0143] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A control method for a photovoltaic grid-connected inverter, characterized in that, Includes the following steps: S1. Collect the voltage and current of the photovoltaic array, the voltage and current of the grid connection point, the temperature of the inverter power devices, and the sequence images of the all-sky optical imaging sensor, and integrate them to generate a multi-dimensional real-time dataset. S2. Based on the sequence image data in the multidimensional real-time dataset, a convolutional long short-term memory network model is used to predict the spatiotemporal characteristics of cloud dynamics and generate a spatiotemporal irradiance prediction map of the photovoltaic array surface. S3. Based on the grid connection point voltage and current data in the multidimensional real-time dataset, the grid impedance spectrum data is identified by injecting a disturbance signal into the inverter output current reference and analyzing the grid response. S4. Call the spatiotemporal irradiance prediction map, and combine the electrical characteristic model of the photovoltaic array with the global optimization algorithm to solve the global maximum power point voltage trajectory that maximizes the total output power of the array in the prediction time domain.

2. The control method for a photovoltaic grid-connected inverter according to claim 1, characterized in that, The multidimensional real-time dataset includes the DC voltage and current values ​​of the photovoltaic array, the instantaneous AC and DC voltage and current values ​​at the grid connection point, the temperature of the power devices, and a sequence of digital images of the sky. The spatiotemporal irradiance prediction map includes the predicted values ​​of solar radiation intensity on each spatial grid within the future time pane. The grid impedance spectrum data includes the grid impedance amplitude and phase angle characteristics within a preset frequency band. The global maximum power point voltage trajectory includes the optimal DC bus reference voltage at each time step within the future time pane.

3. The control method for a photovoltaic grid-connected inverter according to claim 1, characterized in that, The specific steps for generating the multidimensional real-time dataset are as follows: S111, synchronously acquire electrical status data of photovoltaic array, thermal status data of key power devices inside inverter, and sky image sequence data of all-sky optical imaging sensor. S112. Based on the precise time protocol, the electrical status data, thermal status data and sky image sequence data are timestamped and synchronized to ensure the time consistency of the data source; S113. The heterogeneous data after synchronization processing is encapsulated into a unified record format and integrated to generate the multidimensional real-time dataset containing electrical, thermal and environmental conditions.

4. The control method for a photovoltaic grid-connected inverter according to claim 3, characterized in that, The specific steps for generating the spatiotemporal irradiance prediction map are as follows: S211. Perform geometric correction on the sequence image data to eliminate lens distortion, and perform radiometric calibration to establish the mapping relationship between pixel gray values ​​and solar radiation intensity, thus completing image preprocessing. S212. Input the preprocessed image sequence into a pre-trained convolutional long short-term memory network, extract cloud spatial features from its convolutional layers, and learn and model the temporal evolution of the spatial features from its long short-term memory layers. S213. The future predicted image sequence output by the network is converted into quantized irradiance values ​​through the mapping relationship and projected onto the physical coordinate system of the photovoltaic array to form the spatiotemporal irradiance prediction map.

5. The control method for a photovoltaic grid-connected inverter according to claim 4, characterized in that, The specific steps for identifying the power grid impedance spectrum data are as follows: S311. In the inverter's conventional grid-connected current reference command, a preset, non-interference pseudo-random binary sequence is superimposed as a disturbance current reference signal. S312. Synchronously acquire the grid connection point voltage response waveform and current response waveform caused by the disturbance current reference signal, and extract the voltage and current response components caused by the disturbance through bandpass filtering. S313. Perform Fast Fourier Transform on the voltage response component and the current response component respectively, and obtain the grid impedance spectrum data by calculating the complex ratio of the two in the frequency domain.

6. The control method for a photovoltaic grid-connected inverter according to claim 5, characterized in that, The specific steps for solving the global maximum power point voltage trajectory are as follows: S411. For each future time point in the spatiotemporal irradiance prediction map, load the spatial distribution data of the irradiance at that time point into a preset photovoltaic array topology model. S412. Based on the mathematical model of the photovoltaic module, calculate the power-voltage characteristic curve of the entire photovoltaic array under the current irradiance distribution and temperature conditions, and use particle swarm optimization or genetic algorithm to perform a global search on the curve to determine the global maximum power point voltage. S413. Traverse all time points within the predicted time domain and repeat the loading, calculation, and search process to form a sequence composed of the global maximum power point voltage at each time point, thereby obtaining the global maximum power point voltage trajectory.

7. The control method for a photovoltaic grid-connected inverter according to claim 1, characterized in that, The method also includes step S5: S5. Construct a multi-objective model predictive control framework, call the grid impedance spectrum data and the global maximum power point voltage trajectory, integrate multiple performance indicators for online rolling optimization and solve, and generate a set of optimal pulse width modulation drive signals to control the power switching transistors of the inverter. The aforementioned performance indicators include power point tracking error, total harmonic distortion of grid-connected current, DC bus voltage fluctuation, and thermal stress of power devices.

8. The control method for a photovoltaic grid-connected inverter according to claim 7, characterized in that, The execution steps of the multi-objective model predictive control framework are as follows: S511. In each control cycle, the state-space prediction model of the inverter-grid coupling system is updated online based on the latest identified grid impedance spectrum data. S512. Construct a multi-objective cost function composed of the weighted average of the multiple performance indicators, and dynamically adjust the weight coefficients of each performance indicator according to the spatiotemporal irradiance prediction map and the power grid impedance spectrum data. S513. Under the system operation safety constraints, the optimal control sequence that minimizes the sum of the cost functions in the future prediction time domain is solved by numerical optimization algorithm, and its first element is converted into the optimal pulse width modulation drive signal.

9. The control method for a photovoltaic grid-connected inverter according to claim 8, characterized in that, The dynamic adjustment of the weighting coefficients of each performance indicator specifically includes: When the spatiotemporal irradiance prediction map indicates that there will be severe power fluctuations in the future, the weight of the index used to characterize DC bus voltage fluctuations will be increased. When the grid impedance spectrum data shows that the grid exhibits high impedance characteristics in a specific harmonic frequency band, the weight of the index used to characterize the total harmonic distortion of the grid-connected current is increased.

10. A control system for a photovoltaic grid-connected inverter, characterized in that, The system is used to implement the control method for the photovoltaic grid-connected inverter according to any one of claims 1-9, and the system includes: The multidimensional data acquisition module is used to simultaneously acquire photovoltaic array voltage and current, grid voltage and current, device temperature and sky sequence images, and integrate them to generate the multidimensional real-time dataset. The spatiotemporal irradiance prediction module is used to process the sequence image data and generate a spatiotemporal irradiance prediction map of the photovoltaic array surface by learning and extrapolating the dynamic spatiotemporal characteristics of the cloud layer. The grid impedance identification module is used to actively inject small disturbance currents and analyze the grid response to calculate and output the grid impedance spectrum data of the grid connection point in real time. The forward-looking power optimization module is used to calculate and generate the global maximum power point voltage trajectory based on the non-uniform irradiance distribution of the spatiotemporal irradiance prediction map through a global optimization search algorithm. The multi-objective model predictive control module is used to perform rolling time-domain optimization calculations based on a system model that is dynamically updated according to the grid impedance spectrum data, under the constraints of multiple weighted performance indicators, and finally generate and output the optimal pulse width modulation drive signal for controlling the inverter power switching transistors.

Images