Artificial intelligence-based photovoltaic tracking support method
The photovoltaic tracking system, which utilizes distributed sensor networks and lightweight neural networks, solves the problems of response lag and control deviation in existing photovoltaic tracking systems under complex weather conditions, achieving efficient and low-cost optimization of photovoltaic power generation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TIANJIN MINGXUAN NEW ENERGY TECHNOLOGY DEVELOPMENT CO LTD
- Filing Date
- 2026-04-03
- Publication Date
- 2026-06-30
AI Technical Summary
Existing photovoltaic tracking systems suffer from slow response, large control deviations, reliance on high-cost sensors, and a lack of multi-dimensional environmental fusion perception capabilities under complex weather conditions, resulting in power generation efficiency failing to approach the theoretical optimal value.
A distributed low-density sensor network is used to acquire multi-dimensional environmental data. Convolutional neural networks and long short-term memory networks are combined to predict illumination. Tracking angle commands are generated through a lightweight fully connected neural network, and dynamic correction is performed using power feedback and gradient correction mechanisms. Low-latency closed-loop control is achieved by combining servo motor control.
It achieves high-precision tracking under complex lighting conditions, reduces hardware costs, improves power generation efficiency, enhances the system's adaptability and robustness, and increases the average daily power generation by 12.5% to 38% compared to traditional systems.
Smart Images
Figure CN122308472A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of solar photovoltaic power generation technology, specifically relating to a photovoltaic tracking bracket method based on artificial intelligence. Background Technology
[0002] With the continuous advancement of photovoltaic power generation technology, photovoltaic tracking brackets, as key equipment for improving solar energy utilization efficiency, are receiving increasing attention for their intelligent and adaptive control capabilities. In recent years, artificial intelligence technology has been gradually introduced into photovoltaic tracking systems in order to achieve more efficient and accurate solar position tracking and angle adjustment. Traditional fixed photovoltaic systems cannot dynamically adjust their tilt angle according to the sun's trajectory, resulting in significantly limited daily energy capture efficiency. While early mechanical single-axis or dual-axis tracking systems could achieve basic angle adjustment, their control logic largely relied on preset astronomical algorithms or simple light intensity threshold judgments, lacking the ability to perceive and respond to complex environmental disturbances. In cloudy, partially obscured, or rapidly changing weather scenarios, they are prone to misadjustment or lag, making it difficult to maintain optimal irradiance reception.
[0003] Among them, the AI-based photovoltaic tracking bracket method aims to improve the system's adaptability to unsteady lighting conditions by replacing traditional rule-based control strategies with data-driven models. This type of method typically combines sensor data with historical operational information to construct predictive or optimization models to output real-time tracking angle commands. However, existing technologies still suffer from structural deficiencies in model architecture design, environmental perception dimensions, and control closed-loop mechanisms, which limit their deployment efficiency and economic feasibility in practical engineering scenarios.
[0004] In existing technologies, some solutions focus on fault diagnosis rather than active tracking control, achieving anomaly detection only by comparing power deviation and angle error. They lack forward-looking modeling and dynamic adjustment capabilities for the solar trajectory, resulting in the system failing to approach the theoretical maximum power generation efficiency even under normal operating conditions. Another type of solution, while incorporating illumination distribution prediction and support response models for closed-loop optimization, heavily relies on high-density sensor arrays to collect full-domain irradiance data, significantly increasing hardware costs and maintenance complexity. Its time-sharing computing architecture introduces non-negligible processing latency, easily causing tracking phase inaccuracies during the rapidly changing solar altitude angles of dawn and dusk. More critically, the feedback correction mechanism of such methods relies solely on instantaneous power generation indicators, failing to effectively integrate multi-dimensional environmental variables such as cloud movement trends, shadow projection from nearby objects, and atmospheric transparency decay. This causes the model output to deviate from the true optimal solution under complex illumination interference, making it difficult to guarantee system robustness and long-term operational stability. Therefore, there is an urgent need for a photovoltaic tracking support method that integrates lightweight intelligent reasoning, multi-source heterogeneous environmental perception, and low-latency execution mechanisms to overcome the bottlenecks of existing technologies in real-time performance, adaptability, and energy efficiency synergistic optimization. Summary of the Invention
[0005] The purpose of this invention is to provide an artificial intelligence-based photovoltaic tracking bracket method to solve the technical defects of existing photovoltaic tracking systems in complex weather conditions and unsteady lighting environments, such as response lag, large control deviation, reliance on high-cost sensors, and lack of multi-dimensional environmental fusion perception capabilities, which lead to the power generation efficiency failing to approach the theoretical optimal value.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0007] An artificial intelligence-based photovoltaic tracking bracket method includes:
[0008] Step 1: Collect multi-dimensional environmental perception data. The real-time irradiance, sky imaging information, atmospheric temperature and humidity, wind speed and direction, and projection status of nearby obstructions in the area where the photovoltaic array is located are obtained through a distributed low-density sensor network. The irradiance sensor sampling frequency is 10 times per second, the sky imaging device resolution is 2560 pixels by 1440 pixels, and the frame rate is 30 frames per second. All data are uploaded to the edge computing node after being timestamped.
[0009] Step 2: Extract dynamic features of illumination. Based on the continuous time sequence of sky images, use a convolutional neural network to extract cloud morphology evolution features. Combine the irradiance fluctuation curve and use a long short time memory network to model the local illumination change trend. Output the irradiance distribution prediction sequence for the next 30 seconds, with a prediction time step of 1 second.
[0010] Step 3: Generate initial tracking angle command. The irradiance distribution prediction sequence, the current solar zenith angle and azimuth angle, and the atmospheric transparency attenuation coefficient are used as input variables. The pre-trained lightweight fully connected neural network model is input. The model has 3 hidden layers with 128, 64 and 32 neurons in each layer, respectively. The activation function is a modified linear unit. The target values of pitch and rotation angles of the dual-axis tracking support are output. The update period is once every 2 seconds.
[0011] Step 4: Dynamic correction is performed by integrating power feedback. The DC power signal at the output of the photovoltaic module is monitored in real time with a sampling interval of 0.5 seconds. The ratio between the current actual power and the maximum possible power estimated based on the ideal incident angle model is calculated. When the ratio is lower than 95% for three consecutive times, the feedback correction module is triggered to adjust the tracking angle command. The adjustment range is determined according to the gradient ascent algorithm with a step size of 0.5 degrees per iteration cycle.
[0012] Step 5: Execute low-latency collaborative control, transmit the corrected angle command to each partition drive unit via CAN bus. The drive unit has a built-in servo motor controller with a response delay of less than 100 milliseconds. During the execution, the position encoder data is fed back in real time to form a closed-loop control loop, ensuring that the bracket attitude stably tracks the command value and the time for a single angle adjustment does not exceed 3 seconds.
[0013] In step 1, the sky imaging device is installed above the geometric center of the photovoltaic array, with a field of view covering the area from the zenith to the horizon. The lens surface is coated with an anti-fouling and hydrophobic coating. In the image preprocessing stage, noise reduction, white balance correction, and polar coordinate transformation are performed in sequence to enhance the accuracy of cloud edge recognition.
[0014] In step 2, the convolutional neural network contains four convolutional layers and two max-pooling layers. The convolutional kernel sizes are 7x7, 5x5, 3x3, and 3x3, respectively, with 16, 32, 64, and 64 channels. The pooling window is 2x2. The feature map is flattened and then input into the long short-term memory network, which contains two recursive units with a hidden state dimension of 50. The forget gate structure is pruned to reduce the computational load.
[0015] In step 2, the irradiance fluctuation curve and the cloud motion vector are spatiotemporally aligned. The optical flow method is used to estimate the translation speed and direction of the cloud cluster and map it onto the projection area of the photovoltaic array surface. Combined with terrain elevation data, the arrival time of the shadow is predicted, and the angle pre-adjustment mechanism is activated in advance.
[0016] In step 3, the lightweight fully connected neural network model is quantized and compressed before deployment. The weight parameters are converted from 32-bit floating-point to 8-bit fixed-point format, reducing the model size to 25% of the original size and the inference time to less than 15 milliseconds per cycle. It can run stably on embedded ARM architecture processors.
[0017] In step 3, the atmospheric transparency attenuation coefficient is obtained by looking up the real-time aerosol optical thickness and water vapor content in a table. The data source is the local micro-weather station or the regional numerical weather prediction system, and the update frequency is once every 10 minutes.
[0018] In step 4, the maximum possible power estimation model is calculated based on the AM1.5 standard spectrum and the projected area under the current solar incidence angle. The power reduction factor is introduced to consider the component temperature effect, and the temperature coefficient is set to -0.4% / degree Celsius. The measured temperature is sampled from the thermistor integrated on the back panel of the component.
[0019] In step 4, the search direction of the gradient ascent algorithm is determined by the power difference results of the two consecutive steps. If the power increases, the current adjustment direction is maintained; otherwise, the adjustment is reversed. At the same time, a momentum term is introduced to prevent oscillations at local peaks, and the momentum coefficient is fixed at 0.8.
[0020] Preferably, in step 5, the CAN bus communication protocol adopts a custom high-efficiency frame structure. Each control frame carries target angle, check code and timestamp information. The data field length is compressed to within 8 bytes, supporting synchronous control of up to 64 drive units, and the communication conflict rate is less than 0.1%.
[0021] In step 5, the servo motor controller is equipped with an incremental encoder with a resolution of up to 4096 pulses per revolution. Combined with the PID control algorithm, it achieves an angle positioning accuracy of ±0.2 degrees and has the ability to resist wind disturbances without losing steps under gusts of 12 m / s.
[0022] It also includes a historical operation data storage module, which continuously records environmental perception data, control command sequences and power generation performance indicators for no less than one year, forming a structured database for subsequent model retraining and parameter optimization. The data update cycle is once per hour.
[0023] The historical operating data is used to train the next generation of neural network models offline. A transfer learning strategy is adopted to fine-tune the output layer weights for specific seasons or climate patterns while retaining the basic feature extraction capabilities, thereby improving long-term adaptability.
[0024] The method supports multiple photovoltaic tracking brackets to form a cluster for collaborative operation. It shares key environmental perception information and abnormal event alarms through a wireless mesh network, realizes joint response to regional light disturbances, and controls communication latency to within 200 milliseconds.
[0025] Preferably, the method automatically switches to astronomical algorithm-assisted mode during dawn and dusk, and calculates the initial value of the sun's trajectory by combining GPS positioning and precise clock signals, so as to avoid visual perception failure due to insufficient light and ensure continuous tracking capability throughout the day.
[0026] The method integrates a fault self-diagnosis function. By analyzing the deviation sequence between the angle command and the actual feedback, it can determine abnormal states such as mechanical jamming, transmission wear, or sensor drift. When the deviation continues to exceed 1.5 degrees for more than 10 seconds, a maintenance warning is issued.
[0027] Compared with the prior art, the beneficial technical effects of the present invention are as follows:
[0028] This invention overcomes the technical bottlenecks of slow response and insufficient control precision of traditional photovoltaic tracking brackets under complex lighting conditions by constructing a closed-loop control system that integrates multi-source environmental perception and lightweight artificial intelligence reasoning.
[0029] This invention significantly reduces hardware deployment costs and maintenance difficulty through low-density sensor configuration, while the illumination prediction mechanism based on joint modeling of sky images and irradiance time series effectively improves the system's forward-looking perception capability of cloud movement and local occlusion.
[0030] This invention achieves high-speed inference on resource-constrained edge devices through a lightweight neural network model, ensuring real-time updates of control commands.
[0031] By introducing a power feedback and gradient correction mechanism, this invention enables the system to dynamically approach the maximum power point under non-ideal operating conditions, thus avoiding energy loss due to model deviation.
[0032] This invention ensures the stability and anti-interference capability of mechanical execution through high-precision closed-loop control of the servo drive unit. The overall solution achieves tracking accuracy better than ±0.5 degrees and control delay less than 3 seconds without relying on a large-scale sensor network. Daily power generation is increased by at least 12.5% compared to traditional single-axis tracking systems and by up to 38% compared to fixed-tilt installation methods.
[0033] This invention further enhances the system's environmental adaptability at the regional scale through a cluster collaboration mechanism, and the model iteration capability driven by historical data ensures continuous optimization of long-term operational efficiency. This invention achieves the organic unity of high energy efficiency, strong robustness, low cost and intelligent control, and provides a reliable technical path for improving the economic efficiency of large-scale photovoltaic power generation systems. Attached Figure Description
[0034] Figure 1 This is a flowchart illustrating the overall technical architecture of a photovoltaic tracking bracket method based on artificial intelligence proposed in this invention.
[0035] Figure 2 This is a schematic diagram of the core principle framework of the present invention, which integrates multi-source environmental perception and lightweight artificial intelligence reasoning. Detailed Implementation
[0036] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.
[0037] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.
[0038] In the embodiments of the present invention, the same reference numerals denote the same components, and for the sake of brevity, detailed descriptions of the same components are omitted in different embodiments. It should be understood that the thickness, length, width, and other dimensions of various components in the embodiments of the present invention shown in the accompanying drawings, as well as the overall thickness, length, width, and other dimensions of the integrated device, are merely illustrative and should not constitute any limitation on the present invention; the term "multiple" in the present invention refers to two or more (including two).
[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.
[0040] Currently, as photovoltaic (PV) tracking brackets are key equipment for improving solar energy utilization efficiency, their intelligent and adaptive control capabilities are receiving increasing attention. Traditional fixed PV systems cannot dynamically adjust their tilt angle according to the sun's trajectory, resulting in significantly limited daily energy capture efficiency. While early mechanical single-axis or dual-axis tracking systems can achieve basic angle adjustment, their control logic often relies on preset astronomical algorithms or simple light intensity threshold judgments, lacking the ability to perceive and respond to complex environmental disturbances. In cloudy, partially obscured, or rapidly changing weather scenarios, they are prone to misadjustment or lag, making it difficult to maintain optimal irradiance reception. To address these technical problems, this invention proposes to construct a lightweight artificial intelligence inference model, integrate multi-source heterogeneous environmental data input, and establish a low-latency closed-loop feedback mechanism to achieve dynamic and accurate prediction and real-time adjustment of the solar incidence angle. This significantly improves the system's adaptability, robustness, and energy capture efficiency without requiring large-scale sensor arrays, and is applied to an artificial intelligence-based PV tracking bracket method.
[0041] refer to Figure 1 and Figure 2 The overall technical solution architecture and core principle framework of the photovoltaic tracking bracket method based on artificial intelligence proposed in this invention are clearly demonstrated. For example... Figure 1As shown, the system consists of a distributed low-density sensor network, edge computing nodes, a servo drive unit cluster, and a historical data storage module, forming a complete closed loop. Figure 2 As shown, multi-source environmental perception data is fused and then input into a lightweight neural network model to generate initial tracking commands. Dynamic correction is performed through power feedback, and finally, attitude adjustment is completed by a high-precision actuator. The specific steps of the method of this invention are as follows:
[0042] Step 1: Collect multi-dimensional environmental perception data. A distributed, low-density sensor network is used to acquire real-time irradiance, sky imaging information, atmospheric temperature and humidity, wind speed and direction, and the projection status of nearby obstructions in the area where the photovoltaic array is located. Specifically, the distributed low-density sensor network includes several key components: an irradiance sensor continuously collects global irradiance data on the horizontal and inclined planes at a sampling frequency of 10 times per second; a sky imaging device is installed above the geometric center of the photovoltaic array, with a field of view covering the area from the zenith to the horizon, and its lens surface is coated with an anti-fouling and hydrophobic coating, with an image resolution of 2560 pixels multiplied by 1440 pixels and a frame rate of 30 frames per second; a micro-weather station simultaneously collects atmospheric temperature and humidity and wind speed and direction data; the projection status of nearby obstructions is obtained by analyzing the static obstacle outlines in the sky image and their dynamic shadow boundaries on the photovoltaic panel. All sensor data streams are timestamped at the edge computing nodes according to a high-precision hardware clock to ensure strict synchronization of multi-source heterogeneous data in the time dimension, with an alignment error controlled within 1 millisecond, and then uniformly uploaded to the edge computing nodes for subsequent processing.
[0043] Step 2: Extract dynamic features of illumination. Based on a continuous time-series sky image sequence, a convolutional neural network is used to extract cloud morphology evolution features. Combined with the irradiance fluctuation curve, a long short-term memory network is used to model the local illumination change trend, outputting an irradiance distribution prediction sequence for the next 30 seconds, with a prediction time step of 1 second. Specifically, the sky image sequence first undergoes an image preprocessing stage, sequentially performing denoising, white balance correction, and polar coordinate transformation. Denoising uses a nonlocal mean filtering algorithm, white balance correction is based on the gray world assumption, and polar coordinate transformation maps the original Cartesian coordinate system sky image to a polar coordinate plane centered on the zenith, thereby enhancing the accuracy of cloud edge recognition and simplifying the subsequent feature extraction process. The preprocessed image sequence is input into a customized convolutional neural network containing 4 convolutional layers and 2 max-pooling layers. The convolutional kernel sizes are 7x7, 5x5, 3x3, and 3x3, with 16, 32, 64, and 64 channels, respectively. The pooling window is 2x2, and the activation function is a modified linear unit. After flattening, the resulting high-dimensional feature vector is fed into a two-layer recursive unit Long Short-Term Memory (LSTM) network with a hidden state dimension of 50. The forget gate structure is pruned to reduce computational load, thus effectively modeling the long-term dependencies of cloud morphology. Simultaneously, the irradiance fluctuation curve is spatiotemporally aligned with the cloud motion vector extracted from the image sequence. Optical flow is used to estimate the cloud's translational velocity and direction, and this motion vector is mapped onto the projection area of the photovoltaic array surface. Combined with pre-stored terrain elevation data, the shadow arrival time is predicted, and an angle pre-adjustment mechanism is initiated in advance. Finally, the model, which integrates visual features and temporal irradiance data, outputs an irradiance distribution prediction sequence of length 30, where each element represents the expected irradiance intensity in the next second (t=1 to 30).
[0044] Step 3: Generate the initial tracking angle command. The irradiance distribution prediction sequence, current solar zenith angle and azimuth angle, and atmospheric transparency attenuation coefficient are used as input variables. These are input to a pre-trained lightweight fully connected neural network model. This model has three hidden layers with 128, 64, and 32 neurons per layer, respectively. The activation function uses a modified linear unit. The output is the target values for the pitch and rotation angles of the dual-axis tracking support, updated every 2 seconds. Specifically, the current solar zenith angle and azimuth angle are calculated in real-time by a built-in astronomical algorithm module based on a precise clock signal and GPS positioning coordinates, with a calculation accuracy better than 0.1 degrees. The atmospheric transparency attenuation coefficient is obtained from a table based on real-time aerosol optical thickness and water vapor content, sourced from local micro-weather stations or regional numerical weather prediction systems, updated every 10 minutes. All input variables are normalized to... The model is then input into the interval. This lightweight fully connected neural network model undergoes quantization and compression before deployment, converting the weight parameters from 32-bit floating-point to 8-bit fixed-point format. The model size is reduced to 25% of the original size, and the inference time is reduced to less than 15 milliseconds per cycle. It can run stably on embedded ARM architecture processors, meeting real-time requirements.
[0045] Step 4: Dynamic correction is performed using integrated power feedback. The DC power signal at the output of the photovoltaic module is monitored in real time with a sampling interval of 0.5 seconds. The ratio between the current actual power and the maximum possible power estimated based on the ideal incident angle model is calculated. When this ratio is below 95% for three consecutive times, the feedback correction module is triggered to adjust the tracking angle command. The adjustment magnitude is determined based on the gradient ascent algorithm, with a step size of 0.5 degrees per iteration cycle. Specifically, the maximum possible power estimation model is calculated based on the AM1.5 standard spectrum and the projected area under the current solar incident angle, and a power reduction factor is introduced considering the module temperature effect. The temperature coefficient is set to -0.4% / degree Celsius, and the measured temperature is sampled from the thermistor integrated on the module backsheet. The formula for calculating the power ratio is as follows:
[0046]
[0047] in, To measure DC power, Based on predicted irradiance Current support angle and measured temperature Estimated ideal power. When If this state persists for 3 sampling periods (i.e., 1.5 seconds), the system determines that there is a model bias or unforeseen environmental disturbance, and then activates the gradient ascent algorithm. The search direction of this algorithm is determined by the power difference results of two consecutive iterations. If the power increases, the current adjustment direction is maintained; otherwise, the adjustment is reversed. A momentum term is introduced to prevent oscillations at local peaks, and the momentum coefficient is fixed at 0.8. The angle adjustment step size for each iteration is fixed at 0.5 degrees, until... Restore to over 95% or reach the maximum number of iterations limit.
[0048] Step 5: Execute low-latency collaborative control. The corrected angle command is transmitted to each partition drive unit via the CAN bus. The drive unit has a built-in servo motor controller with a response latency of less than 100 milliseconds. During execution, position encoder data is fed back in real time, forming a closed-loop control circuit to ensure that the support posture stably tracks the command value. The time to complete a single angle adjustment does not exceed 3 seconds. Specifically, the CAN bus communication protocol adopts a custom high-efficiency frame structure. Each control frame carries the target angle, check code, and timestamp information. The data field length is compressed to within 8 bytes, supporting synchronous control of up to 64 drive units with a communication conflict rate of less than 0.1%. Each drive unit is equipped with an incremental encoder with a resolution of up to 4096 pulses per revolution. Combined with the PID control algorithm, the angle positioning accuracy reaches ±0.2 degrees, and the wind disturbance resistance meets the requirement of not losing synchronization under gusts of 12 m / s. Position feedback data is transmitted back to the edge computing node at a frequency of 100 times per second for real-time monitoring of the execution status and participation in the correction decision of the next cycle.
[0049] Furthermore, this invention includes a historical operation data storage module, which continuously records environmental perception data, control command sequences, and power generation performance indicators for at least one year, forming a structured database for subsequent model retraining and parameter optimization. The data update cycle is once per hour. Historical operation data is used for offline training of the next-generation neural network model, employing a transfer learning strategy. While retaining the basic feature extraction capability, the output layer weights are fine-tuned for specific seasons or climate patterns to improve long-term adaptability. The method supports a cluster collaborative working mode of multiple photovoltaic tracking brackets, sharing key environmental perception information and abnormal event alarms through a wireless mesh network to achieve joint response to regional light disturbances, with communication latency controlled within 200 milliseconds. The method automatically switches to an astronomical algorithm-assisted mode during dawn and dusk, combining GPS positioning and precise clock signals to calculate the initial value of the solar trajectory, avoiding visual perception failure due to insufficient light and ensuring continuous tracking capability around the clock. The method integrates a fault self-diagnosis function, analyzing the deviation sequence between angle commands and actual feedback to determine abnormal states such as mechanical jamming, transmission wear, or sensor drift. When the deviation persists for more than 1.5 degrees for more than 10 seconds, a maintenance warning is issued.
[0050] To illustrate the technical effects of this invention more specifically, an application example is constructed at a large-scale ground-mounted photovoltaic power station in Northwest China. This power station deploys 500 photovoltaic tracking brackets of this invention, each equipped with a complete low-density sensor network and edge computing unit. On a cloudy summer afternoon, the system successfully detected a rapidly moving cumulus cloud and, using the illumination prediction model in step 2, predicted 8 seconds in advance that it would cast a shadow on the photovoltaic array. The initial command generated in step 3 already included a pre-adjustment action, shifting the bracket angle towards the expected unshaded area. However, due to the complex internal structure of the cloud, the actual degree of shading was slightly higher than predicted, causing the power ratio in step 4 to drop to 92% within 3 seconds. The feedback correction module then activated, fine-tuning the bracket angle by 1.5 degrees through three gradient-up iterations (taking 1.5 seconds), restoring the power ratio to 96.5%. The entire process, from detection to correction completion, took only 4.5 seconds, far superior to the response time of over 10 seconds for traditional systems. According to the statistics of operating data for a month, the daily power generation of the power station increased by 13.2% compared with the adjacent traditional single-axis tracking system, which verifies the superior performance of the present invention under complex weather conditions.
[0051] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A photovoltaic tracking bracket method based on artificial intelligence, characterized in that: The specific steps include the following: Step 1: Collect multi-dimensional environmental perception data. Use a distributed low-density sensor network to obtain real-time irradiance, sky imaging information, atmospheric temperature and humidity, wind speed and direction, and the projection status of nearby obstructions in the area where the photovoltaic array is located. All data are uploaded to the edge computing node after being timestamped. Step 2: Extract dynamic features of illumination. Based on the continuous time sequence of sky images, use a convolutional neural network to extract cloud morphology evolution features. Combine the irradiance fluctuation curve and use a long short time memory network to model the local illumination change trend, and output the irradiance distribution prediction sequence for the next 30 seconds. Step 3: Generate initial tracking angle command. The irradiance distribution prediction sequence, the current solar zenith angle and azimuth angle, and the atmospheric transparency attenuation coefficient are used as input variables. The pre-trained lightweight fully connected neural network model is input, and the activation function is a modified linear unit. The target values of the pitch angle and rotation angle of the dual-axis tracking support are output. The update cycle is once every 2 seconds. Step 4: Dynamic correction is performed by integrating power feedback. The DC power signal at the output of the photovoltaic module is monitored in real time with a sampling interval of 0.5 seconds. The ratio between the current actual power and the maximum possible power estimated based on the ideal incident angle model is calculated. When the ratio is lower than 95% for three consecutive times, the feedback correction module is triggered to adjust the tracking angle command. The adjustment range is determined according to the gradient ascent algorithm with a step size of 0.5 degrees per iteration cycle. Step 5: Execute low-latency collaborative control, transmit the corrected angle command to each zone drive unit via CAN bus. The drive unit has a built-in servo motor controller, and during the execution process, it provides real-time feedback of position encoder data to form a closed-loop control circuit, ensuring that the bracket's attitude stably tracks the command value, and the time for a single angle adjustment does not exceed 3 seconds.
2. The photovoltaic tracking bracket method based on artificial intelligence according to claim 1, characterized in that: In step 1, the sky imaging device is installed above the geometric center of the photovoltaic array, with a field of view covering the area from the zenith to the horizon. The lens surface is coated with an anti-fouling and hydrophobic coating. In the image preprocessing stage, noise reduction, white balance correction, and polar coordinate transformation are performed sequentially to enhance the accuracy of cloud edge recognition. The irradiance sensor in the sensor has a sampling frequency of 10 times per second, and the sky imaging device has a resolution of 2560 pixels multiplied by 1440 pixels and a frame rate of 30 frames per second.
3. The photovoltaic tracking bracket method based on artificial intelligence according to claim 1, characterized in that: In step 2, the convolutional neural network contains four convolutional layers and two max-pooling layers. The convolutional kernel sizes are 7x7, 5x5, 3x3, and 3x3, respectively, with 16, 32, 64, and 64 channels. The pooling window is 2x2. The feature map is flattened and then input into a long short-term memory network, which contains two recursive units. The hidden state dimension is 50, and the forget gate structure is pruned to reduce the computational load. The prediction time step is 1 second.
4. The photovoltaic tracking bracket method based on artificial intelligence according to claim 1, characterized in that: In step 2, the irradiance fluctuation curve and the cloud motion vector are spatiotemporally aligned. The optical flow method is used to estimate the translation speed and direction of the cloud cluster and map it onto the projection area of the photovoltaic array surface. Combined with terrain elevation data, the arrival time of the shadow is predicted, and the angle pre-adjustment mechanism is activated in advance.
5. The photovoltaic tracking bracket method based on artificial intelligence according to claim 1, characterized in that: In step 3, the lightweight fully connected neural network model is quantized and compressed before deployment. The weight parameters are converted from 32-bit floating-point to 8-bit fixed-point format, reducing the model size to 25% of the original size and reducing the inference time to less than 15 milliseconds per cycle, enabling it to run stably on embedded ARM architecture processors. The fully connected neural network model has 3 hidden layers, with 128, 64 and 32 neurons in each layer, respectively.
6. The photovoltaic tracking bracket method based on artificial intelligence according to claim 1, characterized in that: In step 3, the atmospheric transparency attenuation coefficient is obtained by looking up the real-time aerosol optical thickness and water vapor content in a table. The data source is the local micro-weather station or the regional numerical weather prediction system, and the update frequency is once every 10 minutes.
7. The photovoltaic tracking bracket method based on artificial intelligence according to claim 1, characterized in that: In step 4, the maximum possible power estimation model is calculated based on the AM1.5 standard spectrum and the projected area under the current solar incidence angle. The power reduction factor is introduced to consider the component temperature effect. The temperature coefficient is set to -0.4% / degree Celsius. The measured temperature is sampled from the thermistor integrated on the back panel of the component. The DC power signal sampling interval is 0.5 seconds.
8. The photovoltaic tracking bracket method based on artificial intelligence according to claim 1, characterized in that: In step 4, the search direction of the gradient ascent algorithm is determined by the power difference results of the two consecutive steps. If the power increases, the current adjustment direction is maintained; otherwise, the adjustment is reversed. At the same time, a momentum term is introduced to prevent oscillations at local peaks, and the momentum coefficient is fixed at 0.
8.
9. The photovoltaic tracking bracket method based on artificial intelligence according to claim 1, characterized in that: In step 5, the CAN bus communication protocol adopts a custom high-efficiency frame structure. Each control frame carries the target angle, check code, and timestamp information. The data field length is compressed to within 8 bytes, supporting synchronous control of up to 64 drive units, with a communication conflict rate of less than 0.1%. The servo motor controller is equipped with an incremental encoder with a resolution of up to 4096 pulses per revolution. Combined with the PID control algorithm, it achieves an angle positioning accuracy of ±0.2 degrees and meets the requirements of not losing synchronization under gusts of 12 m / s. The response delay of the servo motor controller is less than 100 milliseconds.
10. The photovoltaic tracking bracket method based on artificial intelligence according to claim 1, characterized in that: It also includes a historical operation data storage module, which continuously records environmental perception data, control command sequences, and power generation performance indicators for no less than one year, forming a structured database for subsequent model retraining and parameter optimization. The data update cycle is once per hour. The method supports a cluster collaborative working mode of multiple photovoltaic tracking brackets, sharing key environmental perception information and abnormal event alarms through a wireless mesh network, with communication latency controlled within 200 milliseconds. It automatically switches to an astronomical algorithm-assisted mode during dawn and dusk, combining GPS positioning and precise clock signals to calculate the initial value of the solar trajectory. It integrates a fault self-diagnosis function, which analyzes the deviation sequence between angle commands and actual feedback, and issues a maintenance warning when the deviation lasts for more than 1.5 degrees for more than 10 seconds.