A method and system for maximum power point tracking control of photovoltaic modules based on AI

Abstract

This invention, entitled "An AI-Based Maximum Power Point Tracking (GMPP) Control Method and System for Photovoltaic Modules," belongs to the field of photovoltaic power generation control technology. The technical problem it aims to solve is the insufficient global optimization capability of existing GMPP technologies under complex operating conditions, reliance on environmental sensors, and lack of adaptive tracking mechanisms under sudden changes in illumination. The key technical solutions are: collecting electrical parameters of the photovoltaic module to generate P-V curves; determining the initial GMPP range using an improved quantum ocean predation algorithm; achieving sensorless MPPT inference at the edge using a lightweight multilayer perceptron and outputting fine-tuning values ​​for control parameters; predicting illumination and temperature trends in the cloud using a long short-term memory network to adjust the algorithm optimization strategy and model parameters in advance; real-time monitoring of sudden changes in illumination to achieve mode switching between algorithm re-optimization and model fine-tuning; and designing a system including modules for photovoltaic data acquisition and feature processing to adapt to the execution of this method.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power generation control technology, specifically relating to an AI-based method and system for maximum power point tracking control of photovoltaic modules. Background Technology

[0002] With the global energy transition accelerating, photovoltaic (PV) power generation, as a core form of clean and renewable energy utilization, faces an increasingly urgent need for efficiency optimization. Maximum power point tracking (MPPT) technology, a key component in improving PV system efficiency, directly determines the energy capture capability of PV modules under complex operating conditions. Currently, PV systems often face complex conditions such as sudden changes in illumination, localized shading, and temperature fluctuations, causing the power-voltage (PV) characteristic curve of PV modules to exhibit multi-peak characteristics. Traditional MPPT algorithms and existing intelligent optimization and AI-integrated MPPT solutions all have significant technical bottlenecks, making it difficult to simultaneously achieve a balance between global optimization, dynamic response, and steady-state accuracy. They also cannot simultaneously meet the core operational requirements of PV systems, such as sensorless operation, adaptation to complex operating conditions, edge deployment, and rapid dynamic response.

[0003] Relevant patent documents retrieved:

[0004] The document, published in China (CN118939073A), discloses a method and system for MPPT photovoltaic power optimization based on photovoltaic modules. This method uses a deep learning model to predict changes in illumination and temperature, and combines edge computing to optimize power tracking points.

[0005] The document, published in China (CN119311074A), discloses a method, device, equipment, and medium for MPPT control based on a photovoltaic array. It obtains the power point tracking stability coefficient and historical tracking data through a deep learning model, determines the voltage and current reference values, and realizes MPPT control.

[0006] The prior art represented by the aforementioned documents has at least the following unresolved technical problems or defects: (1) Existing patented technical solutions only process local data, lack global optimization capabilities, cannot effectively cope with multi-peak PV curves under local shadows, and lack a fast mode switching mechanism when illumination changes suddenly, resulting in poor dynamic adaptability. The relevant evidence is that the solution does not design a global optimization algorithm architecture and mode switching logic for illumination changes, and only relies on edge computing to perform local optimization of power tracking points.

[0007] (2) Existing patented technology solutions rely on light and temperature sensors to collect environmental data and do not achieve sensorless MPPT inference. The relevant evidence is that the solution uses light and temperature data collected by sensors combined with historical tracking data to train the model and determine reference values. There is no implicit mapping design between electrical parameters and environmental factors, and additional sensor hardware is required.

[0008] (3) The existing patented technical solutions do not construct a collaborative mechanism of global optimization and local fine-tuning. The dynamic response speed is difficult to meet the needs of the working conditions of rapid changes in illumination. The relevant evidence is that the solution only uses a deep learning model to fit historical data to determine the reference values ​​of voltage and current, without a control architecture of global coarse positioning + local fine adjustment.

[0009] (4) Existing quantum heuristic algorithm-based MPPT solutions (such as methods based on Grover's algorithm and quantum particle swarm optimization) can improve global optimization capabilities, but they do not combine lightweight AI models to achieve local fine-tuning and are not optimized for edge computing power constraints. They have high computational complexity and are not suitable for embedded deployment. The relevant evidence is that such solutions only use a single quantum algorithm for optimization, without local fine-tuning design of lightweight AI models, and the algorithm has not been optimized for computing power and energy consumption in an engineering manner, making it unsuitable for embedded hardware such as MCU / FPGA.

[0010] (5) None of the existing MPPT technology solutions have built a collaborative architecture of cloud prediction and edge control, and cannot achieve forward-looking parameter adjustment and control. The relevant evidence is that the existing solutions either rely solely on local computing at the edge or rely solely on deep learning for simple environmental prediction, without cloud-based trend prediction based on time series data or reverse correction logic for edge algorithms and model parameters.

[0011] In view of this, the present invention is hereby proposed. Summary of the Invention

[0012] To address the aforementioned technical problems in existing technologies, this invention provides an AI-based maximum power point tracking (MPPT) control method and system for photovoltaic modules. This system solves the problems of insufficient global optimization capability, reliance on environmental sensors, lack of forward-looking parameter adjustment, and lack of adaptive tracking mechanism under sudden changes in illumination when existing photovoltaic modules perform MPPT under complex operating conditions.

[0013] To achieve the above objectives, the technical solution of the present invention is as follows: A photovoltaic module maximum power point tracking control method based on AI, comprising: S1. Collect voltage, current, and power data of the photovoltaic module, and generate power-voltage curve data of the photovoltaic module; S2. Based on the power-voltage curve data, the preliminary range of the global maximum power point is determined by using an improved quantum ocean predation algorithm that incorporates the optimization characteristics of quantum particle swarm optimization. S3. Based on the preliminary range of the global maximum power point, at the edge end, combined with the output voltage and current of the photovoltaic module, a sensorless maximum power point tracking inference is performed through a lightweight multilayer sensor, and the fine-tuning amount of the control parameters is output. S4. Based on the fine-tuning amount of the control parameters, combined with the electrical parameters and energy consumption data of the photovoltaic module, the changing trends of light and temperature are predicted in the cloud through a long short-term memory network. The optimization strategy of the improved quantum ocean predation algorithm and the fine-tuning parameters of the lightweight multilayer sensor are adjusted in advance according to the changing trends. S5. Control the photovoltaic module based on the fine-tuning amount of the control parameters, and monitor the sudden change in the illumination state of the photovoltaic module. If a sudden change in illumination is detected, the improved quantum ocean predation algorithm is triggered to perform global optimization again to update the initial range of the global maximum power point. If the illumination remains stable, the fine-tuning mode of the lightweight multilayer sensor is maintained and the fine-tuning amount of the control parameters is continuously output.

[0014] Furthermore, in step S2, the execution process of the improved quantum ocean predation algorithm is as follows: Initial population positions are randomly generated on the power-voltage curve. Calculate the power at each location as the fitness value. ; based on and The population is iterated through an ocean predation algorithm to search for the global optimal position; during the population iteration, the position of the particles is nonlinearly transformed through a quantum rotation gate; In the later stages of the algorithm, quantum collapse is introduced for localized mining; When the number of iterations reaches the maximum value or the power change in multiple consecutive iterations is less than the set power change threshold, the algorithm terminates and outputs the current global optimal position and its neighborhood as the preliminary range of the global maximum power point.

[0015] Furthermore, the nonlinear transformation of the particle's position through a quantum rotation gate specifically refers to: using a quantum rotation gate... The quantum rotating gate performs a nonlinear transformation on the particle position. The matrix is:

[0016] When the particle moves away from the global maximum power point Smaller It has a relatively large size and a wide search range; when the particle approaches the global maximum power point... Approaching 0, fine-tuning is then performed.

[0017] in, To act on the first A quantum rotation gate matrix for individual particles. The rotation angle determines the magnitude of the change in the search direction. The initial rotation angle is... For adjustment coefficients, The fitness is the global optimum of the current population. Combining the position update of the quantum rotation gate, the updated position of the quantum state evolution is obtained by the following formula:

[0018]

[0019] in, , These represent the probability amplitudes of the corresponding qubits. For the first Sub-particles, For the first The probability that a particle is in its ground state. , These represent the upper and lower boundaries of the photovoltaic optimization quantity.

[0020] Furthermore, the localized mining that introduces quantum collapse specifically refers to: Calculate the first The average historical best position of the population is determined, and particles collapse towards the central region of the global optimum. The specific formula is as follows:

[0021]

[0022] in, For the first The average historical best position of the population across generations, The total number of individuals, For the first The particle up to the [number]th The optimal historical position of an individual in a given era. For the first The global optimal position of the generation. The quantum collapse contraction-expansion coefficient, For the log-normal distribution term, The plus or minus sign is randomly selected with a 50% probability.

[0023] Furthermore, in step S3, the preliminary range of the global maximum power point is specifically as follows: The collected output voltage and current of the photovoltaic module are normalized and characterized to eliminate the influence of dimensions, and the voltage change is calculated. With change in current Meanwhile, by combining the initial range of the global maximum power point to narrow down the input variable range, a lightweight feature set is obtained; Based on historical current, voltage and corresponding change data, combined with ambient light and temperature information, a sensorless lightweight multilayer perceptron model suitable for edge computing is trained. The lightweight feature set is input into the lightweight multilayer perceptron model, and the output is the fine-tuning amount of the control parameters.

[0024] Furthermore, the lightweight multilayer perceptron model adopts a three-layer structure consisting of an input layer, a hidden layer, and an output layer. The hidden layer uses the ReLU activation function, and the output layer undergoes a linear transformation, as shown in the following formula:

[0025]

[0026] in, For the first The output value of each hidden layer neuron For the input layer The first feature is passed to the hidden layer. The weights of each neuron, For the input layer A normalized feature, For the bias term of the hidden layer, The number of input layer features. For fine-tuning the output control parameters, For the hidden layer The nth neuron to the output layer The weights of each neuron, This represents the number of neurons in the hidden layer. For output layer bias terms; The lightweight multilayer perceptron model is converted from 32-bit floating-point type to 8-bit integer type through quantization perception training. After the normalized fine-tuning quantity of its output is mapped back to the actual physical control quantity range, it is clamped within a preset range and combined with the incremental conductance method to compensate for model error.

[0027] Furthermore, in step S4, the process of predicting the changing trends of light and temperature in the cloud through a long short-term memory network, and adjusting the optimization strategy of the improved quantum ocean predation algorithm and the fine-tuning parameters of the lightweight multilayer perceptron in advance, is as follows: According to the preset sampling period, the control parameter fine-tuning amount, electrical parameters and energy consumption data of the edge end are synchronously collected and timestamped to construct a time series dataset with equal time steps; After normalizing the time series dataset and denoising it using a sliding window, a time series feature set is obtained; the time series feature set is then input into a cloud-based long short-term memory network to output predicted values ​​for illumination and temperature. Based on the predicted trends of light and temperature, the optimization strategy of the improved quantum ocean predation algorithm and the fine-tuning parameters of the lightweight multilayer perceptron are adjusted in advance. The cloud-based system sends the adjusted and improved quantum ocean predation algorithm optimization parameter set and the multilayer perceptron fine-tuning parameter set to the edge control unit via a high-speed communication link, completing the real-time parameter update.

[0028] Furthermore, the optimization strategy of the improved quantum ocean predation algorithm is adjusted in advance based on the predicted trends of light and temperature, specifically as follows: Based on the predicted trends of light and temperature, the quantum rotation angle of the improved quantum ocean predation algorithm is corrected in reverse. Global iterative contraction-expansion coefficient Population search step size Quantum collapse probability ; Based on the positional drift trend of the global maximum power point, the search weight of the corresponding voltage range is adjusted; based on the abrupt change trend of illumination, the global exploration capability or local convergence accuracy of the algorithm is adjusted.

[0029] Furthermore, the fine-tuning parameters of the lightweight multilayer perceptron are adjusted in advance as follows: The changes in light and temperature predicted by the cloud-based Long Short-Term Memory network are transformed into dimensionless trend factors. As the input signal for adjusting the parameters of the multilayer perceptron, it changes the model's sensitivity to input voltage and current. The formula is:

[0030] in, The adjusted new output weight matrix, The baseline weight matrix trained for the multilayer perceptron model. , Adjust the gain for the weights.

[0031] Furthermore, the monitoring of sudden changes in light intensity specifically includes: Based on the collected real-time parameters and historical cached parameters, the power and current change rates are calculated. According to the mapping relationship between power, current change rates, and illumination changes, a comprehensive abrupt change index is calculated, using the following formula:

[0032] in, The comprehensive quantitative indicators used for judgment, , The rate of change of power and current. For power weighting coefficients, This is the current weighting coefficient; The comprehensive mutation index is compared with the preset change rate threshold to quantitatively determine whether a light mutation has occurred. If a light mutation is determined, the lightweight multilayer sensor fine-tuning mode is immediately terminated, and the improved quantum ocean predation algorithm is triggered to re-optimize globally based on the latest collected power-voltage curve data, lock the preliminary range of the global maximum power point under the new operating condition, and update the baseline value of the control parameters.

[0033] Secondly, an AI-based photovoltaic module maximum power point tracking control system, applied to the AI-based photovoltaic module maximum power point tracking control method described above, includes: Photovoltaic data acquisition module: used to acquire voltage, current and power data of photovoltaic modules, generate power-voltage curve data of photovoltaic modules, and acquire output voltage, current and electrical parameters and energy consumption data of photovoltaic modules. Feature processing module: It is used to receive and process various data collected by the photovoltaic acquisition module, transmit the processed power-voltage curve data to the global optimal judgment module, transmit the processed output voltage and current to the decision fusion module, and transmit the processed control parameter fine-tuning amount, electrical parameters, and energy consumption data to the decision fusion module. The decision fusion module is used to perform sensorless maximum power point tracking inference using a lightweight multilayer perceptron based on the preliminary range of the global maximum power point and the output voltage and current of the photovoltaic module, and output fine-tuning values ​​for the control parameters. It is also used to predict the trends of light and temperature changes using a long short-term memory network based on the fine-tuning values ​​of the control parameters and the electrical parameters and energy consumption data of the photovoltaic module, thereby adjusting the optimization strategy of the improved quantum ocean predation algorithm and the fine-tuning parameters of the lightweight multilayer perceptron in advance. Simultaneously, it is used to monitor sudden changes in light intensity; if a sudden change in light intensity is detected, the global optimal judgment module is triggered to re-optimize globally; if the light intensity remains stable, the fine-tuning mode of the lightweight multilayer perceptron is maintained. Global Optimization Module: Based on the power-voltage curve data, it determines the preliminary range of the global maximum power point by using an improved quantum ocean predation algorithm that incorporates the optimization characteristics of quantum particle swarm optimization. It can also respond to trigger commands to perform global optimization again through the algorithm. Output module: Used to output control signals based on the fine-tuning amount of the control parameters to control the photovoltaic module.

[0034] The beneficial effects of this invention are as follows: (1) Enhance global optimization capability and adapt to complex working conditions: The improved quantum ocean predation algorithm combines quantum rotating gate nonlinear transformation and quantum collapse local mining mechanism, enabling particles to jump out of local extreme points to the global optimal GMPP. It can effectively handle the multi-peak characteristics of PV curves under local shadow, greatly reduce the probability of getting trapped in local optimal solutions, and ensure accurate locking of the global maximum power point under complex working conditions. (2) Achieve sensorless control and improve system robustness: By learning the implicit mapping relationship between current, voltage and environmental factors, the dependence on light and temperature sensors is eliminated. This not only reduces the hardware cost of the photovoltaic control system and simplifies the installation and maintenance process, but also reduces the risk of system failure caused by sensor failure, and significantly improves the robustness of system operation. (3) Achieve forward-looking control and improve dynamic response speed: Relying on the cloud LSTM network, based on the historical electrical parameters, energy consumption data and control fine-tuning of the edge, the trend of light and temperature changes can be accurately predicted, and the optimization strategy of QMPA and the fine-tuning parameters of MLP at the edge can be adjusted in advance, so that the control system has the ability to predict environmental changes, greatly improves the dynamic response speed under sudden or rapid changes in light, and effectively reduces the power loss of photovoltaic system. (4) Adaptable to edge embedded deployment and has high engineering application value: The MLP model adopts a minimalist network structure design, combined with ReLU activation function and quantization perception training technology, which significantly reduces the computational complexity and memory occupation of the model, enabling the AI ​​algorithm to run in real time on low-cost embedded hardware such as MCU / FPGA, with both control performance and engineering practicality, and has a wide range of promotion prospects. Attached Figure Description

[0035] Figure 1 A flowchart of an AI-based maximum power point tracking control method for photovoltaic modules provided in an embodiment of the present invention; Figure 2 An architecture diagram of an AI-based photovoltaic module maximum power point tracking control system provided in an embodiment of the present invention; Figure 3 This is an architecture diagram of the improved quantum ocean predation algorithm QMPA provided in an embodiment of the present invention; Figure 4 This is an architecture diagram of a lightweight multilayer perceptron (MLP) provided in an embodiment of the present invention. Figure 5 Example graph of PV characteristic curve provided in the embodiment of the present invention; Figure 6 This is a data diagram of the method under sudden changes in light intensity provided in an embodiment of the present invention; Figure 7 This is a comparison chart of performance test data under standard unobstructed steady-state conditions provided in an embodiment of the present invention; Figure 8 This is a data comparison chart under different light intensities for a local shading multi-peak condition provided in an embodiment of the present invention; Figure 9 This is a comparison chart of performance test data under sudden change in light intensity provided in an embodiment of the present invention. Detailed Implementation

[0036] The technical solution of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0037] It should be noted that, unless otherwise specifically stated, the relative arrangement and numerical expressions of the components and steps described in these embodiments should not be construed as limiting the scope of the invention.

[0038] The following description of exemplary embodiments is merely illustrative and is not intended to limit the invention or its application or use in any way. Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail herein, but where applicable, such techniques, methods, and apparatus should be considered part of this specification.

[0039] Terminology Explanation: MPPT: Maximum Power Point Tracking, a core technology in the field of photovoltaic power generation control, is used to track the maximum power output point of photovoltaic modules under different operating conditions to improve power generation efficiency; GMPP: Global Maximum Power Point, the globally optimal power output point in the power-voltage curve of a photovoltaic module, as opposed to a local maximum power point; QMPA: An improved quantum ocean predation algorithm that incorporates the characteristics of quantum particle swarm optimization. QPSO: Quantum Particle Swarm Optimization, a probabilistic optimization algorithm whose characteristics are incorporated into QMPA to enhance the algorithm's global optimization capability; MLP: Multilayer Perceptron, a lightweight model trained with quantized perception and weight pruning, deployed at the edge to achieve sensorless MPPT inference; LSTM: Long Short-Term Memory Network, a time-series feature learning network deployed in the cloud, used to capture environmental time-series patterns and predict trends in light and temperature changes; PV curve: Power-voltage curve, which reflects the relationship between the output power and output voltage of a photovoltaic module, and is the basic data carrier for the MPPT algorithm optimization. INC: Incremental conductance method, used to compensate for inference errors in lightweight MLP models, verify and correct control parameter fine-tuning, and improve the steady-state accuracy of MPPT control; MCU: Microcontroller Unit, an embedded hardware core at the edge, used to deploy lightweight algorithms and models to achieve real-time data acquisition and control of photovoltaic modules; FPGA: Field Programmable Gate Array, embedded hardware adapted to edge computing power and energy consumption constraints; ADC: Analog-to-Digital Converter, a sampling module integrated into the edge control unit, converts the analog voltage and current signals of photovoltaic modules into digital signals for algorithm processing; DC-DC: A power electronic device that uses the control parameters output by the present invention to adjust the operating point of a photovoltaic module and achieve maximum power point tracking.

[0040] Example 1 See Figure 1 , Figure 1 This is a flowchart of an AI-based maximum power point tracking control method for photovoltaic modules proposed in this invention. Specific steps may include: S1. Collect voltage, current, and power data of photovoltaic modules and generate power-voltage (PV) curve data of photovoltaic modules; S2. Based on PV curve data, the preliminary range of the Global Maximum Power Point (GMPP) is determined by incorporating the improved quantum ocean predation algorithm (QMPA) with quantum particle swarm optimization characteristics. See Figure 3 The execution process of the improved quantum ocean predation algorithm is as follows: S21. Randomly generate initial population positions on the power-voltage curve. Calculate the power at each location as the fitness value. The specific formula is as follows:

[0041] in, For the first The position of each individual element represents its voltage value. , Represents the minimum and maximum voltage. To generate random numbers, For the first The fitness value of each particle. The real-time voltage and current are sampled at the corresponding locations; S22, based on and The population is iterated using an ocean predation algorithm to search for the Global Optimal Position (GMPP). During population iteration, a quantum rotation gate is used to perform a nonlinear transformation on the particle positions. Specifically, the nonlinear transformation on the particle positions using the quantum rotation gate is as follows: S221, Through the quantum rotating door Nonlinear transformation of particle position, quantum rotating gate The matrix is:

[0042] S222. When the particle is far from the global maximum power point... Smaller It has a relatively large size and a wide search range; when the particle approaches the global maximum power point... Approaching 0, fine-tuning is then performed.

[0043] in, To act on the first A quantum rotation gate matrix for individual particles. The rotation angle determines the magnitude of the change in the search direction. The initial rotation angle is... For adjustment coefficients, The fitness is the global optimum of the current population. S223. Combining the position update of the quantum rotation gate, the updated position of the quantum state evolution is obtained, using the following formula:

[0044]

[0045] in, , These represent the probability amplitudes of the corresponding qubits. For the first Sub-particles, For the first The probability that a particle is in its ground state. , These represent the upper and lower boundaries of the photovoltaic optimization quantity.

[0046] S23. In the later stages of the algorithm, quantum collapse is introduced for local mining. The specific steps of local mining include: S231, Calculate the first The average historical best position of the population is determined, and particles collapse towards the central region of the global optimum. The specific formula is as follows:

[0047]

[0048] in, For the first The average historical best position of the population across generations, The total number of individuals, For the first The particle up to the [number]th The optimal historical position of an individual in a given era. For the first The global optimal position of the generation. The quantum collapse contraction-expansion coefficient, For the log-normal distribution term, To simulate the random collapse of a quantum state by randomly selecting a plus or minus sign with a 50% probability.

[0049] S24. When the number of iterations reaches the maximum value or the power change in multiple consecutive iterations is less than the set power change threshold, the algorithm terminates and outputs the current global optimal position and its neighborhood as the preliminary range of the global maximum power point.

[0050] S3. Based on the initial range of GMPP, the output voltage and current of the photovoltaic module are combined at the edge end, and sensorless maximum power point tracking (MPPT) inference is performed through a lightweight multilayer sensor (MLP) to output the fine-tuning amount of the control parameters. Specifically, the preliminary range of the global maximum power point is as follows: S31. Normalize and characterize the collected output voltage and current of the photovoltaic module to eliminate the influence of dimensions, and calculate the voltage change. With change in current Meanwhile, by combining the initial range of the global maximum power point to narrow down the input variable range, a lightweight feature set is obtained; S32. Based on historical current, voltage and corresponding change data, combined with ambient light and temperature information, train a sensorless lightweight multilayer perceptron model suitable for edge devices. S33. Input the lightweight feature set into the lightweight MLP model and output the fine-tuning amount of the control parameters; See Figure 4 The training is suitable for lightweight MLP models at the edge. Specifically, during training, the corresponding relationship between temperature, light, current, and voltage is learned. The changes in current and voltage are mapped to the corresponding changes in temperature and light environment, reducing the input of temperature and light data and retaining only current and voltage data to achieve sensorless MPPT inference. The lightweight multilayer perceptron model adopts a three-layer structure consisting of an input layer, a hidden layer, and an output layer. The hidden layer uses the ReLU activation function, and the output layer undergoes a linear transformation. The specific formula is as follows:

[0051]

[0052] in, For the first The output value of each hidden layer neuron For the input layer The first feature is passed to the hidden layer. The weights of each neuron, For the input layer A normalized feature, For the bias term of the hidden layer, The number of input layer features. For fine-tuning the output control parameters, For the hidden layer The nth neuron to the output layer The weights of each neuron, This represents the number of neurons in the hidden layer. For output layer bias terms; The lightweight multilayer perceptron model converts 32-bit floating-point data to 8-bit integer data through quantization perceptual training, reducing computational complexity and energy consumption, and making it suitable for embedded environments such as MCUs / FPGAs; the specific formula is as follows:

[0053] in, The weights are quantized 8-bit integers. For the original 32-bit floating-point weights, For the number of quantization bits, This is the quantization scaling factor. This is a rounding function; The normalized fine-tuning value of the MLP output is mapped back to the actual physical control range to obtain the fine-tuning value of the control parameters. The rationality of the fine-tuning amount is verified in real time. If it exceeds the preset range, clamping is performed. At the same time, the incremental conductance method (INC) is used to compensate for model errors to ensure steady-state accuracy. The specific formula is as follows:

[0054] in, The fine-tuning amount for the final output control parameters. This is the fine-tuning amount for the actual physical control quantity.

[0055] S4. Based on the fine-tuning amount of the control parameters, combined with the electrical parameters and energy consumption data of the photovoltaic module, the changing trends of light and temperature are predicted in the cloud using LSTM, and the optimization strategy of the improved quantum ocean predation algorithm and the fine-tuning parameters of the lightweight multilayer sensor are adjusted in advance. Specifically, the process of predicting the changing trends of light and temperature using LSTM in the cloud, and adjusting the optimization strategy of the improved quantum ocean predation algorithm and the fine-tuning parameters of the lightweight multilayer perceptron in advance, is as follows: S41. According to the preset sampling period, the control parameter fine-tuning amount, electrical parameters and energy consumption data at the edge end are synchronously collected and timestamped to construct a time series dataset with equal time steps; S42. After normalizing the data and denoising the time series dataset using a sliding window, a time series feature set is obtained; the time series feature set is input into a cloud LSTM to capture the time series features of the environment and illumination, and output the predicted values ​​of illumination and temperature. S43. Based on the predicted trends of light and temperature, adjust the optimization strategy of the improved quantum ocean predation algorithm and the fine-tuning parameters of the lightweight multilayer perceptron in advance. Based on the predicted trends of illumination and temperature, optimization strategies for fine-tuning the parameters of QMPA and lightweight MLP are adjusted in advance. After all predictions and parameter adjustments are completed in the cloud, the optimized parameter set of the improved QMPA and the fine-tuning parameter set of the lightweight MLP are sent to the local edge control unit via a high-speed communication link. Upon receiving the instruction, the local control unit updates the algorithm parameters in real time. The specific adjustment steps are as follows: S431. Based on the predicted trends of light and temperature, the GMPP position of photovoltaics will drift. According to the changing trends of light and temperature output by the cloud LSTM network, the core optimization parameters and strategies of the improved QMPA are corrected in reverse, and the quantum rotation angle is adjusted. Global iterative contraction-expansion coefficient Population search step size Quantum collapse probability ; S432. When the light intensity increases or the temperature decreases, the GMPP drifts towards the high-voltage side, and the QMPA expands the search weight of the high-voltage region in advance; when the light intensity decreases or the temperature increases, the GMPP drifts towards the low-voltage side, and the QMPA expands the search weight of the low-voltage region in advance; when the light intensity changes abruptly, the QMPA improves the global exploration capability in advance; when the light intensity is stable, the QMPA improves the local convergence accuracy in advance. The formula for adjusting the quantum rotation angle is as follows:

[0056] in, The adjusted quantum rotation angle, As the reference rotation angle, , The gain coefficient is adjusted for illumination and temperature. , For predicted trends in light and temperature; The formula for adjusting the global iterative contraction-expansion coefficient is as follows:

[0057] in, These are the adjusted coefficients. , The maximum and minimum thresholds for the global iterative contraction-expansion coefficient are... , These are the thresholds for drastic changes in light intensity and temperature, respectively. The formula for adjusting the population search step size is:

[0058] in, The adjusted search step size, As the reference step size, This is the step size amplification factor; The formula for adjusting the quantum collapse probability is:

[0059] in, This is the adjusted quantum collapse probability. , These represent the high collapse probability during stability and the low collapse probability during exploration, respectively. For light The sensitivity to mutations at that time.

[0060] S433. Convert the changes in illumination and temperature predicted by the cloud-based LSTM into dimensionless trend factors. As the input signal for adjusting the parameters of the multilayer perceptron, it changes the model's sensitivity to input voltage and current. The formula is:

[0061] in, The adjusted new output weight matrix, The baseline weight matrix trained for the multilayer perceptron model. , Adjust the gain for the weights.

[0062] S44. The cloud sends the adjusted and improved quantum ocean predation algorithm optimization parameter set and the multilayer perceptron fine-tuning parameter set to the edge control unit through a high-speed communication link to complete the real-time parameter update.

[0063] S5. The photovoltaic module is controlled based on the fine-tuning of the control parameters, and the sudden change in light intensity is monitored simultaneously. If a sudden change in light intensity is detected, the improved quantum ocean predation algorithm is triggered to perform global optimization again. If the light intensity remains stable, the fine-tuning mode of the lightweight multilayer sensor is maintained.

[0064] Specifically, the steps for monitoring sudden changes in light intensity include: S51. Based on the collected real-time parameters and historical cached parameters, calculate the power and current change rates. Based on the mapping relationship between power, current change rates, and illumination changes, calculate the comprehensive abrupt change index, using the following formula:

[0065] in, The comprehensive quantitative indicators used for judgment, , The rate of change of power and current. For power weighting coefficients, This is the current weighting coefficient; S52. Compare the comprehensive mutation index with the preset change rate threshold to quantitatively determine whether a light mutation has occurred. If it is determined to be a light mutation, immediately terminate the lightweight MLP fine-tuning mode, trigger QMPA to re-optimize the global system based on the latest collected power-voltage curve data, lock the preliminary range of the global maximum power point under the new operating condition, and update the control parameter baseline value. If the illumination is determined to be stable, the lightweight MLP fine-tuning mode is maintained continuously. Based on the real-time collected data, the inference results of the MLP model are continuously updated, and the fine-tuning amount of the control parameters is dynamically adjusted to achieve accurate MPP tracking. At the same time, the historical parameter cache is updated synchronously.

[0066] This embodiment details an AI-based maximum power point tracking (GMPP) control method for photovoltaic modules, constructing a collaborative control architecture of edge control and cloud prediction. By integrating the characteristics of quantum particle swarm optimization (QMPA) with quantum rotating gate nonlinear transformation and quantum collapse mechanisms, a preliminary range for GMPP is determined. Based on this preliminary range, a lightweight microprocessor-based logic flow (MLP) is deployed at the edge, using only voltage and current data as input to complete sensorless inference and output fine-tuning of control parameters. Simultaneously, an LSTM is built in the cloud to predict the changing trends of illumination and temperature, and based on this, the optimization strategy of QMPA and the weight parameters of MLP are corrected in advance, achieving look-ahead control of the system. Furthermore, the system monitors the illumination status in real time; when illumination changes abruptly, QMPA is triggered to re-optimize globally, while when illumination is stable, the local fine-tuning mode of MLP is maintained.

[0067] Example 2 See Figure 2 , Figure 2 The system of the maximum power point tracking control method for photovoltaic modules based on AI proposed in this invention may specifically include: Photovoltaic data acquisition module: used to acquire voltage, current and power data of photovoltaic modules, generate power-voltage curve data of photovoltaic modules, and acquire output voltage, current and electrical parameters and energy consumption data of photovoltaic modules. Feature processing module: It is used to receive and process various data collected by the photovoltaic acquisition module, transmit the processed power-voltage curve data to the global optimal judgment module, transmit the processed output voltage and current to the decision fusion module, and transmit the processed control parameter fine-tuning amount, electrical parameters, and energy consumption data to the decision fusion module. The decision fusion module is used to perform sensorless maximum power point tracking inference using a lightweight multilayer perceptron based on the preliminary range of the global maximum power point and the output voltage and current of the photovoltaic module, and output fine-tuning values ​​for the control parameters. It is also used to predict the trends of light and temperature changes using a long short-term memory network based on the fine-tuning values ​​of the control parameters and the electrical parameters and energy consumption data of the photovoltaic module, thereby adjusting the optimization strategy of the improved quantum ocean predation algorithm and the fine-tuning parameters of the lightweight multilayer perceptron in advance. Simultaneously, it is used to monitor sudden changes in light intensity; if a sudden change in light intensity is detected, the global optimal judgment module is triggered to re-optimize globally; if the light intensity remains stable, the fine-tuning mode of the lightweight multilayer perceptron is maintained. Global Optimization Module: Based on the power-voltage curve data, it determines the preliminary range of the global maximum power point by using an improved quantum ocean predation algorithm that incorporates the optimization characteristics of quantum particle swarm optimization. It can also respond to trigger commands to perform global optimization again through the algorithm. Output module: Used to output control signals based on the fine-tuning amount of the control parameters to control the photovoltaic module.

[0068] Example 3 This embodiment details an experiment using this method to perform maximum power point tracking control of photovoltaic modules in a 10kW distributed photovoltaic system. The system is adapted to 250W polycrystalline silicon photovoltaic modules with an open-circuit voltage of 36.8V, a short-circuit current of 8.9A, and a rated power of 250W. The main control parameter is the duty cycle. The specific implementation process is as follows: The STM32H743VI MCU is used as the core control unit at the edge, and a high-precision ADC sampling module is integrated to achieve this. The sampling period is adjusted to simultaneously collect photovoltaic module output voltage, current and converter energy consumption data; an LSTM network training and inference environment is deployed in the cloud, and high-speed data interaction is achieved with the edge through 5G industrial modules, with a communication latency of ≤10ms.

[0069] Preset the initial rotation angle of the improved QMPA adjustment coefficient Global iterative contraction-expansion coefficient The baseline value is 0.8, and the population size is... Maximum number of iterations: 50; power change threshold: 0.2%; GMPP initial range neighborhood. ; The preset lightweight MLP input layer features are 4-dimensional, including... The hidden layer has 16 neurons, the output layer is 1D, the quantization precision is 8 bits, and the weight pruning threshold is [not specified]. The fine-tuning reference threshold is ±0.05, and the incremental conductivity compensation threshold is 0.3%. Preset cloud-based LSTM input timing step size: 30, hidden layer neurons: 48, prediction time: Normalized range [0,1], sliding window width Threshold for comprehensive index of sudden changes in light intensity Power weighting coefficient Current weighting coefficient .

[0070] Based on real-time voltage Current Calculate real-time power; after continuously collecting 100 sets of data, generate the following: Figure 5 The PV characteristic curve example shown is uploaded to the edge algorithm module after the data is denoised by a sliding window.

[0071] See Figure 5 Taking the control round corresponding to the PV characteristic curve shown as an example, 20 initial particle positions are randomly generated within the voltage range [18V, 36.8V] of the PV curve. Calculate the power corresponding to each position as the fitness value. ; Based on particle position and fitness value, an iterative ocean predation algorithm is executed. In each iteration, the particle position is updated through a quantum rotation gate. When the particle moves away from the GMPP (Gross-Gate Predation Probe), ... The movement increased to Expand the search range; when the particle approaches GMPP, Approaching 0, perform a more refined search; In the later stages of the algorithm, around round 30, a quantum collapse mechanism is introduced to calculate the average optimal position and guide particles to collapse towards the global optimal center region. The global iterative contraction-expansion coefficient is then used. Adjusted to 0.9 to improve convergence speed; The algorithm terminates and outputs the globally optimal voltage when the power change is ≤0.2% for 5 consecutive iterations. The corresponding initial GMPP range is [28.2V, 32.2V], and the initial duty cycle reference value is calibrated simultaneously.

[0072] The collected voltage and current data are normalized, the changes between adjacent cycles are calculated, and the input range is narrowed by combining the initial range of GMPP to generate a lightweight feature set. The feature set is input into the quantized lightweight MLP model, and after ReLU activation in the hidden layer and linear transformation in the output layer, the normalized duty cycle fine-tuning is obtained. After inverse normalization, it is mapped to the actual fine-tuning amount. After clamping, it was confirmed to be within ±0.05. Incremental conductance method is used for auxiliary correction to ensure that the power oscillation amplitude after fine-tuning is ≤0.3% and the output target duty cycle is achieved. This drives the DC-DC converter to adjust its operating point.

[0073] edge end press Periodically, it synchronously uploads control parameter fine-tuning amounts, electrical parameters, and energy consumption data (converter losses). ,efficiency After the cloud timestamps are labeled, a 30-step time series dataset is constructed.

[0074] The time-series dataset is normalized and denoised using a sliding window method. The data is then input into an LSTM network to capture temporal features, based on the current... Predicting light intensity and temperature of 27.6℃ for the next 3 minutes, indicating an increase in light intensity and temperature; QMPA parameters are adjusted based on predicted trends; quantum rotation angle from Adjust to Global iterative contraction-expansion coefficient Adjust to 0.75, population search step size Increase by 20%, adjust the quantum collapse probability to 0.3, and expand the search weight of the high-voltage region in advance; MLP parameters adjusted to The duty cycle fine-tuning threshold is relaxed to ±0.06, improving the sensitivity to changes in illumination. The adjusted parameters are sent to the edge via 5G for real-time updates.

[0075] Real-time power acquisition at the edge Historical power Current Historical Current Calculate the rate of change of power Current change rate ; Calculate the comprehensive mutation index. If it is less than the threshold of 5%, the illumination is considered stable, and the MLP fine-tuning mode is maintained. Simulating rapid cloud cover, causing a sudden drop in sunlight to 400 W / m², the calculation yielded... If the threshold is exceeded, and after confirmation for three consecutive sampling cycles, a sudden change in illumination is determined. MLP fine-tuning is immediately terminated, triggering the improved QMPA to re-optimize. Within 5ms, a new initial GMPP range [26.5V, 30.5V] is locked, and the duty cycle reference value is updated. Then switch back to MLP fine-tuning mode to obtain the following result: Figure 6 The data graph of this method under sudden changes in illumination is shown in the figure. As can be seen from the figure, since QMPA has completed the coarse positioning, MLP only needs to make minor adjustments to stabilize the operating point at around 26.5V and the output power at 86.5W, achieving rapid recovery without overshoot or oscillation.

[0076] This embodiment describes in detail the experiment of using the method to perform maximum power point tracking control of photovoltaic modules in a 10kW distributed photovoltaic system. The experiment design fully demonstrates that the method has the characteristics of high-precision global optimization, fast dynamic response and low energy consumption. It can effectively adapt to stable and sudden light conditions, and does not rely on additional sensors, thus having a wide range of application value.

[0077] Example 4 This embodiment details the construction of a photovoltaic MPPT performance testing platform under a controlled laboratory environment. It simulates three typical photovoltaic operating scenarios: standard atmospheric conditions, partial shading, and sudden changes in illumination. The core performance indicators of the deep learning photovoltaic performance prediction method DLEC-MPPT and the photovoltaic array performance learning method DLRS-MPPT, which are discussed in this method and its technical background, are quantitatively tested and compared. The specific process is as follows: Based on the actual engineering needs of photovoltaic MPPT control and the technical shortcomings of existing methods, four core quantitative performance indicators were selected as the evaluation criteria for this comparative experiment, including global optimization capability, dynamic response characteristics, tracking accuracy, and system practicality.

[0078] DLEC-MPPT uses deep learning models to predict changes in illumination and temperature, and combines edge computing to optimize power tracking points; DLRS-MPPT uses a deep learning model to fit the mapping relationship between the power tracking stability coefficient and historical data, and determines the voltage and current reference values ​​to complete MPPT control.

[0079] The four core performance indicators of this method, DLEC-MPPT, and DLRS-MPPT were tested in three scenarios: standard unshaded operation, partial shading with multiple peaks, and sudden change in illumination. In the standard unshaded operation, the illumination intensity was 1000 W / m² ± 50 W / m², the ambient temperature was 25℃ ± 2℃, there was no shading, the photovoltaic PV curve showed a single-peak characteristic, and the system was in a stable operating state. The data obtained after testing each method are shown in Table 1 below. Table 1

[0080] According to Table 1 and Figure 7As shown in the performance test data comparison chart under the standard unobstructed steady-state condition, under the standard unobstructed steady-state condition, this method outperforms the two comparison methods in all core indicators. The global MPP tracking accuracy reaches 99.8%, the energy capture efficiency reaches 99.5%, the steady-state power fluctuation rate is only 0.8%, and the power tracking response time is 45ms. It also has the advantage of being sensorless, which not only ensures high accuracy and high energy utilization in steady-state operation, but also demonstrates excellent operational stability and engineering practicality.

[0081] Under the condition of partial shading and multi-peak operation, the illuminance is distributed in zones from 800W / m² to 1200W / m², forming a multi-peak PV curve with three power peaks. The area of ​​partial shading accounts for 30% of the total area of ​​the photovoltaic module, and the ambient temperature is 25℃±2℃. The data obtained from the tests of the three methods are shown in Table 2 below: Table 2

[0082] According to Table 2 and Figure 8 The comparison chart of test data of various methods under different light intensities in the case of multi-peak PV curves formed by local shading shows that this method has a strong global optimization capability for the case of multi-peak PV curves formed by local shading. The global MPP tracking accuracy reaches 99.6%, the local optimum trap rate is 0, and the global optimization time is only 124.3ms with a steady-state power fluctuation rate of 0.9%. It effectively solves the defects of the comparison method that is prone to getting trapped in local optima, has low global optimization accuracy and long time consumption. It can still stably and efficiently track the global maximum power point in the shadow occlusion scene.

[0083] The initial illumination intensity under the sudden change in illumination condition was 1000 W / m², and the ambient temperature was 25℃±2℃. Two scenarios were simulated: a sudden drop in illumination (1000 W / m²→650 W / m², abrupt change time 0.3 s) and a sudden increase in illumination (650 W / m²→1000 W / m², abrupt change time 0.3 s). The average value of the tests for the two scenarios was taken, and the data obtained after testing each method are shown in Table 3 below: Table 3

[0084] According to Table 3 and Figure 9 As shown in the comparison chart of performance test data under sudden illumination changes, this method significantly outperforms other methods in dynamic response performance under sudden illumination changes, with a response time of 82.7ms, a steady-state recovery time of 152.2ms, an MPP tracking accuracy of 99.5% after the change, and a power overshoot of only 1.2%. Compared with the comparison method, this method significantly shortens the response and steady-state recovery time and reduces the power overshoot. With its rapid illumination change identification and mode switching mechanism, it demonstrates better dynamic adaptability and operational stability, and can better adapt to conditions with rapid illumination changes.

[0085] This embodiment details the comparative tests conducted on the proposed method and existing methods in the technical background under a controlled laboratory environment. The overall test results show that the proposed method is significantly superior to DLEC-MPPT and DLRS-MPPT in key indicators such as global MPP tracking accuracy, steady-state operation stability, global optimization capability, dynamic response speed to sudden changes in illumination, and energy utilization efficiency. It not only solves the defects of existing methods such as insufficient global optimization, susceptibility to local optima, and poor dynamic adaptability, but also has the engineering practicality of sensorless operation. It can achieve efficient and stable MPPT control under various photovoltaic conditions, and its comprehensive performance advantages are obvious.

[0086] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-based photovoltaic module maximum power point tracking control method, characterized by, include: S1. Collect voltage, current, and power data of the photovoltaic module, and generate power-voltage curve data of the photovoltaic module; S2. Based on the power-voltage curve data, the preliminary range of the global maximum power point is determined by using an improved quantum ocean predation algorithm that incorporates the optimization characteristics of quantum particle swarm optimization. S3. Based on the preliminary range of the global maximum power point, at the edge end, combined with the output voltage and current of the photovoltaic module, a sensorless maximum power point tracking inference is performed through a lightweight multilayer sensor, and the fine-tuning amount of the control parameters is output. S4. Based on the fine-tuning amount of the control parameters, combined with the electrical parameters and energy consumption data of the photovoltaic module, the changing trends of light and temperature are predicted in the cloud through a long short-term memory network. The optimization strategy of the improved quantum ocean predation algorithm and the fine-tuning parameters of the lightweight multilayer sensor are adjusted in advance according to the changing trends. S5. Control the photovoltaic module based on the fine-tuning amount of the control parameters, and monitor the sudden change in the illumination state of the photovoltaic module. If a sudden change in illumination is detected, the improved quantum ocean predation algorithm is triggered to perform global optimization again to update the initial range of the global maximum power point. If the illumination remains stable, the fine-tuning mode of the lightweight multilayer sensor is maintained and the fine-tuning amount of the control parameters is continuously output.

2. The AI-based photovoltaic module maximum power point tracking control method of claim 1, wherein, In step S2, the execution process of the improved quantum ocean predation algorithm is as follows: Initial population positions are randomly generated on the power-voltage curve. Calculate the power at each location as the fitness value. ; based on and The population is iterated through an ocean predation algorithm to search for the global optimal position; during the population iteration, the position of the particles is nonlinearly transformed through a quantum rotation gate; In the later stages of the algorithm, quantum collapse is introduced for localized mining; When the number of iterations reaches the maximum value or the power change in multiple consecutive iterations is less than the set power change threshold, the algorithm terminates and outputs the current global optimal position and its neighborhood as the preliminary range of the global maximum power point.

3. The AI-based maximum power point tracking control method for photovoltaic modules according to claim 2, characterized in that, The nonlinear transformation of particle positions via quantum rotation gates specifically involves: using quantum rotation gates... The quantum rotating gate performs a nonlinear transformation on the particle position. The matrix is: When the particle moves away from the global maximum power point Smaller It has a relatively large size and a wide search range; when the particle approaches the global maximum power point... Approaching 0, fine-tuning is then performed. in, To act on the first A quantum rotation gate matrix for individual particles. The rotation angle determines the magnitude of the change in the search direction. The initial rotation angle is... For adjustment coefficients, The fitness is the global optimum of the current population. Combining the position update of the quantum rotation gate, the updated position of the quantum state evolution is obtained by the following formula: in, , These represent the probability amplitudes of the corresponding qubits. For the first Sub-particles, For the first The probability that a particle is in its ground state. , These represent the upper and lower boundaries of the photovoltaic optimization quantity.

4. The AI-based maximum power point tracking control method for photovoltaic modules according to claim 2, characterized in that, The localized mining that introduces quantum collapse specifically refers to: Calculate the first The average historical best position of the population is determined, and particles collapse towards the central region of the global optimum. The specific formula is as follows: in, For the first The average historical best position of the population across generations, The total number of individuals, For the first The particle up to the [number]th The optimal historical position of an individual in a given era. For the first The global optimal position of the generation. The quantum collapse contraction-expansion coefficient, For the log-normal distribution term, The plus or minus sign is randomly selected with a 50% probability.

5. The AI-based maximum power point tracking control method for photovoltaic modules according to claim 1, characterized in that, In step S3, the preliminary range of the global maximum power point is specifically as follows: The collected output voltage and current of the photovoltaic module are normalized and characterized to eliminate the influence of dimensions, and the voltage change is calculated. With change in current Meanwhile, by combining the initial range of the global maximum power point to narrow down the input variable range, a lightweight feature set is obtained; Based on historical current, voltage and corresponding change data, combined with ambient light and temperature information, a sensorless lightweight multilayer perceptron model suitable for edge computing is trained. The lightweight feature set is input into the lightweight multilayer perceptron model, and the output is the fine-tuning amount of the control parameters.

6. The AI-based maximum power point tracking control method for photovoltaic modules according to claim 5, characterized in that, The lightweight multilayer perceptron model adopts a three-layer structure consisting of an input layer, a hidden layer, and an output layer. The hidden layer uses the ReLU activation function, and the output layer undergoes a linear transformation, as shown in the following formula: in, For the first The output value of each hidden layer neuron For the input layer The first feature is passed to the hidden layer. The weights of each neuron, For the input layer A normalized feature, For the bias term of the hidden layer, The number of input layer features. For fine-tuning the output control parameters, For the hidden layer The nth neuron to the output layer The weights of each neuron, This represents the number of neurons in the hidden layer. For output layer bias terms; The lightweight multilayer perceptron model is converted from 32-bit floating-point type to 8-bit integer type through quantization perception training. After the normalized fine-tuning quantity of its output is mapped back to the actual physical control quantity range, it is clamped within a preset range and combined with the incremental conductance method to compensate for model error.

7. The AI-based maximum power point tracking control method for photovoltaic modules according to claim 1, characterized in that, In step S4, the process of predicting the changing trends of light and temperature in the cloud through a long short-term memory network, and adjusting the optimization strategy of the improved quantum ocean predation algorithm and the fine-tuning parameters of the lightweight multilayer perceptron in advance is as follows: According to the preset sampling period, the control parameter fine-tuning amount, electrical parameters and energy consumption data of the edge end are synchronously collected and timestamped to construct a time series dataset with equal time steps; After normalizing the data and denoising it using a sliding window, a time series feature set is obtained. The time-series feature set is input into a cloud-based long short-term memory network, which outputs predicted values ​​for illumination and temperature. Based on the predicted trends of light and temperature, the optimization strategy of the improved quantum ocean predation algorithm and the fine-tuning parameters of the lightweight multilayer perceptron are adjusted in advance. The cloud-based system sends the adjusted and improved quantum ocean predation algorithm optimization parameter set and the multilayer perceptron fine-tuning parameter set to the edge control unit via a high-speed communication link, completing the real-time parameter update.

8. The AI-based maximum power point tracking control method for photovoltaic modules according to claim 7, characterized in that, The optimization strategy of the improved quantum ocean predation algorithm, based on the predicted trends of light and temperature, is adjusted in advance as follows: Based on the predicted trends of light and temperature, the quantum rotation angle of the improved quantum ocean predation algorithm is corrected in reverse. Global iterative contraction-expansion coefficient Population search step size Quantum collapse probability ; Based on the positional drift trend of the global maximum power point, the search weight of the corresponding voltage range is adjusted; based on the abrupt change trend of illumination, the global exploration capability or local convergence accuracy of the algorithm is adjusted.

9. The AI-based maximum power point tracking control method for photovoltaic modules according to claim 7, characterized in that, The fine-tuning parameters of the lightweight multilayer sensor are adjusted in advance as follows: The changes in light and temperature predicted by the cloud-based Long Short-Term Memory network are transformed into dimensionless trend factors. As the input signal for adjusting the parameters of the multilayer perceptron, it changes the model's sensitivity to input voltage and current. The formula is: in, The adjusted new output weight matrix, The baseline weight matrix trained for the multilayer perceptron model. , Adjust the gain for the weights.

10. The AI-based maximum power point tracking control method for photovoltaic modules according to claim 1, characterized in that, The monitoring of sudden changes in light intensity specifically includes: Based on the collected real-time parameters and historical cached parameters, the power and current change rates are calculated. According to the mapping relationship between power, current change rates, and illumination changes, a comprehensive abrupt change index is calculated, using the following formula: in, The comprehensive quantitative indicators used for judgment, , The rate of change of power and current. For power weighting coefficients, This is the current weighting coefficient; The comprehensive mutation index is compared with the preset change rate threshold to quantitatively determine whether a light mutation has occurred. If a light mutation is determined, the lightweight multilayer sensor fine-tuning mode is immediately terminated, and the improved quantum ocean predation algorithm is triggered to re-optimize globally based on the latest collected power-voltage curve data, lock the preliminary range of the global maximum power point under the new operating condition, and update the baseline value of the control parameters.

11. An AI-based maximum power point tracking control system for photovoltaic modules, characterized in that, The method for maximum power point tracking control of a photovoltaic module based on any one of claims 1-10 includes: Photovoltaic data acquisition module: used to acquire voltage, current and power data of photovoltaic modules, generate power-voltage curve data of photovoltaic modules, and acquire output voltage, current and electrical parameters and energy consumption data of photovoltaic modules. Feature processing module: It is used to receive and process various data collected by the photovoltaic acquisition module, transmit the processed power-voltage curve data to the global optimal judgment module, transmit the processed output voltage and current to the decision fusion module, and transmit the processed control parameter fine-tuning amount, electrical parameters, and energy consumption data to the decision fusion module. The decision fusion module is used to perform sensorless maximum power point tracking inference using a lightweight multilayer perceptron based on the preliminary range of the global maximum power point and the output voltage and current of the photovoltaic module, and output fine-tuning values ​​for the control parameters. It is also used to predict the trends of light and temperature changes using a long short-term memory network based on the fine-tuning values ​​of the control parameters and the electrical parameters and energy consumption data of the photovoltaic module, thereby adjusting the optimization strategy of the improved quantum ocean predation algorithm and the fine-tuning parameters of the lightweight multilayer perceptron in advance. Simultaneously, it is used to monitor sudden changes in light intensity; if a sudden change in light intensity is detected, the global optimal judgment module is triggered to re-optimize globally; if the light intensity remains stable, the fine-tuning mode of the lightweight multilayer perceptron is maintained. Global Optimization Module: Based on the power-voltage curve data, it determines the preliminary range of the global maximum power point by using an improved quantum ocean predation algorithm that incorporates the optimization characteristics of quantum particle swarm optimization. It can also respond to trigger commands to perform global optimization again through the algorithm. Output module: Used to output control signals based on the fine-tuning amount of the control parameters to control the photovoltaic module.

Images