Maximum power tracking control method for photovoltaic power generation system

Abstract

The invention discloses a photovoltaic power generation system maximum power tracking control method, and relates to the technical field of power system operation control, and the method comprises the steps: collecting operation data of a photovoltaic power generation system, and carrying out the preprocessing of the operation data, and obtaining feature data, so as to construct a maximum power point prediction model; training a maximum power point prediction model based on the obtained historical operation data sample, and predicting a target working point by using the trained maximum power point prediction model; and in combination with the target working point and an ant colony optimization algorithm, determining an optimal working point, and obtaining control parameters of maximum power tracking by constructing an adaptive voltage controller. According to the MPPT dynamic prediction method based on deep learning, historical operation data and environmental parameters are fused through the LSTM neural network, tracking lag under the conditions of illumination abrupt change and local shadow is avoided, and the prediction precision and response speed of the maximum power point are improved.

Description

Technical Field

[0001] This invention relates to the field of power system operation control technology, and more specifically, to a maximum power point tracking control method for a photovoltaic power generation system. Background Technology

[0002] With the widespread application of renewable energy, photovoltaic (PV) power generation systems have developed rapidly worldwide. To improve the power generation efficiency of PV systems, Maximum Power Point Tracking (MPPT) technology is widely used in the control process of PV arrays. Currently, traditional MPPT methods mainly include Perturband Observe (P&O) and Incremental Conductance (INC). These algorithms are simple in structure and easy to implement, but in practical applications, especially under non-ideal environmental conditions such as rapid changes in light intensity and partial shading of the PV array, they are prone to problems such as slow tracking speed, large power oscillations, and getting trapped in local extrema, making it difficult to ensure that the system always operates at the optimal power point.

[0003] To improve the dynamic performance and stability of traditional methods, some studies have introduced advanced control strategies such as Fuzzy Logic Control (FLC) and Sliding Mode Control (SMC). These methods improve the system's disturbance rejection capability and response speed to some extent, but they still suffer from high implementation complexity, difficulty in parameter tuning, and the challenge of achieving a good balance between steady-state accuracy and dynamic performance. Furthermore, existing MPPT control strategies mostly rely on instantaneous electrical parameters such as the current voltage and current of the photovoltaic system for control decisions, neglecting the mining and utilization of historical operating data of the photovoltaic array and environmental change trends. This makes it difficult for the system to achieve efficient and stable maximum power point tracking under complex operating conditions such as non-uniform irradiation.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] To address the problems in related technologies, this invention proposes a maximum power point tracking control method for photovoltaic power generation systems, thereby overcoming the aforementioned technical problems existing in existing related technologies.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows:

[0007] A method for maximum power point tracking control of a photovoltaic power generation system, the method comprising:

[0008] S1. Collect the operating data of the photovoltaic power generation system and preprocess the operating data to obtain feature data in order to construct the maximum power point prediction model;

[0009] S2. Train the maximum power point prediction model based on the acquired historical operating data samples, and use the trained maximum power point prediction model to predict the target operating point.

[0010] S3. Combining the target operating point with the ant colony optimization algorithm, determine the optimal operating point, and obtain the control parameters for maximum power point tracking by constructing an adaptive voltage controller.

[0011] Furthermore, the operation data of the photovoltaic power generation system is collected and preprocessed to obtain feature data, which is used to construct a maximum power point prediction model, including:

[0012] S11. Set up a data acquisition node between the photovoltaic array and the converter, and collect the operating data of the photovoltaic power generation system in real time through a sensor network;

[0013] S12. Record the running data according to the preset sampling period and establish a time series dataset;

[0014] S13. Apply wavelet transform to perform multi-scale analysis on the established time series dataset, and normalize the analysis results to obtain feature data.

[0015] S14. Based on the obtained feature data, construct a deep learning-based maximum power point prediction model.

[0016] Furthermore, based on the obtained feature data, a deep learning-based maximum power point prediction model is constructed, including:

[0017] S141. Based on the temporal and nonlinear characteristics of the feature data, establish a deep learning framework based on long short-term memory networks.

[0018] S142. Embed a spatial and temporal dual-path attention mechanism in the input layer of the deep learning framework to perform weighted processing on feature data in order to capture the spatial dependencies between sensors and the evolution trend of time series.

[0019] S143. Using the weighted feature data as input and the operating point corresponding to the maximum power point as output, construct a maximum power point prediction model based on deep learning.

[0020] Furthermore, a dual-path attention mechanism, combining spatial and temporal approaches, is embedded in the input layer of the deep learning framework to weight the feature data, thereby capturing the spatial dependencies between sensors and the evolution trends of the time series, including:

[0021] S1421. Based on each time step of the feature data, the spatial attention mechanism is used to calculate the spatial attention weight of the feature data of each sensor in order to capture the spatial dependency between different sensors at the same time.

[0022] S1422. Based on the time series of feature data, the feature data at each time node are weighted using a time attention mechanism to obtain the time attention weight, so as to capture the evolution trend of the time series.

[0023] S1423. Combine spatial attention weights and temporal attention weights to perform joint weighting on the feature data and obtain the weighted feature data.

[0024] Furthermore, the maximum power point prediction model is trained based on the acquired historical operating data samples, and the trained maximum power point prediction model is used to predict the target operating point, including:

[0025] S21. Based on the physical characteristics of photovoltaic cells and environmental factors, establish a mechanism model for maximum power point tracking of photovoltaic power generation system to characterize the nonlinear relationship between output power and voltage.

[0026] S22. Based on the established mechanism model, obtain historical operation data samples of the photovoltaic power generation system, and establish training and validation sets;

[0027] S23. Select the mean square error as the loss function, and train the maximum power point prediction model by combining the established training set and validation set.

[0028] S24. Input the weighted feature data into the trained maximum power point prediction model to predict the target operating point corresponding to the maximum power point.

[0029] Furthermore, the expression for the mechanistic model is:

[0030]

[0031] In the formula, I represents the output current of the photovoltaic cell; I PH Represents the photogenerated current; I0 ​​represents the diode reverse saturation current; q represents the electron charge; V represents the photovoltaic cell terminal voltage; R S A represents the series resistance; A represents the diode ideality factor; K represents the Boltzmann constant; T represents the absolute temperature.

[0032] Furthermore, by combining the target operating point with the ant colony optimization algorithm, the optimal operating point is determined, and the control parameters for maximum power point tracking are obtained by constructing an adaptive voltage controller, including:

[0033] S31. Initialize the individual positions and parameters in the ant colony algorithm, construct the objective function, and use the target working point as the local search center to guide the ant colony to search;

[0034] S32. Based on the search results from the ant colony, dynamically optimize the parameters and update the pheromone concentration of the search path according to the deviation between the target working point and the actual working point.

[0035] S33. Based on the pheromone concentration update results, the search path of individual ants is corrected through a feedback mechanism to obtain the global optimal solution and the optimal working point;

[0036] S34. Based on the optimal operating point, construct an adaptive voltage controller to obtain the control parameters for maximum power point tracking.

[0037] Furthermore, the individual positions and parameters in the ant colony algorithm are initialized, the objective function is constructed, and the target working point is used as the local search center to guide the ant colony in the search, including:

[0038] S311. Initialize the individual positions and parameters in the ant colony algorithm, and perturb the initial parameters using Gaussian noise;

[0039] S312. Based on the parameter distribution after perturbation, the difference between the predicted operating point and the actual operating point of the photovoltaic power generation system is used as the optimization objective to construct an objective function;

[0040] S313. Use the predicted working point as the local search center to guide the ant colony to perform a local search within the search space, and combine evolutionary strategies to optimize the search process.

[0041] Furthermore, based on the optimal operating point, an adaptive voltage controller is constructed to obtain the control parameters for maximum power point tracking, including:

[0042] S341. Based on the obtained optimal operating point, design an adaptive voltage controller that integrates deep reinforcement learning and meta-learning frameworks to generate duty cycle control signals for driving the converter.

[0043] S342. The generated duty cycle control signal is applied to the converter in real time to dynamically adjust the converter's duty cycle to control the output voltage. The control process is optimized by combining the jitter suppression mechanism and the anti-integral saturation mechanism to obtain the control parameters for maximum power point tracking.

[0044] Furthermore, based on the obtained optimal operating point, an adaptive voltage controller integrating deep reinforcement learning and meta-learning frameworks is designed to generate duty cycle control signals for driving the converter, including:

[0045] S3411. Construct a reinforcement learning control strategy based on a deep Q-network framework, combine the obtained optimal operating point, design a reward function with power increment as the objective, and iteratively update the reinforcement learning control strategy through the Q-learning algorithm.

[0046] S3412. Based on the meta-learning mechanism, using the model-independent meta-learning framework, the initial parameters of the strategy suitable for different environmental changes are learned through multi-task training in various photovoltaic power generation scenarios.

[0047] S3413. Integrate the initial parameters of the strategy into the updated reinforcement learning control strategy to generate a control strategy, and output the duty cycle control signal for driving the converter based on the control strategy.

[0048] The beneficial effects of this invention are as follows:

[0049] 1. This invention uses a deep learning-based MPPT dynamic prediction method, which integrates historical operating data and environmental parameters through an LSTM neural network, to avoid tracking lag under conditions of sudden changes in illumination and local shadows, thereby improving the prediction accuracy and response speed of the maximum power point.

[0050] 2. This invention uses an adaptive control parameter optimization method to dynamically adjust the step size and convergence threshold based on the prediction results, thereby avoiding large power oscillations in complex operating conditions caused by fixed parameter control strategies and achieving a balance between steady-state accuracy and dynamic performance.

[0051] 3. This invention constructs a prediction-control collaborative optimization framework, which dynamically integrates deep learning prediction models with mechanistic models. Through confidence assessment and error feedback mechanisms, it avoids the problem of unstable prediction by a single model under abnormal operating conditions, thereby improving the robustness and reliability of the system under all operating conditions.

[0052] 4. This invention optimizes the real-time performance of complex algorithms on resource-constrained devices by combining a lightweight edge deployment scheme with model compression and hardware acceleration technologies. It can achieve millisecond-level response on typical photovoltaic controllers (such as the STM32 series) while maintaining high energy efficiency.

[0053] 5. This invention employs a multi-objective optimization adaptive control strategy, dynamically adjusting control parameters while simultaneously considering indicators such as tracking speed, power quality, and equipment lifespan. This avoids a single optimization objective and improves the overall performance of the photovoltaic system. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart of a maximum power point tracking control method for a photovoltaic power generation system according to an embodiment of the present invention;

[0056] Figure 2 This is a graph of the IV characteristic curve of a photovoltaic cell in a maximum power point tracking control method for a photovoltaic power generation system according to an embodiment of the present invention;

[0057] Figure 3 This is a structural diagram of an LSTM unit in a maximum power point tracking control method for a photovoltaic power generation system according to an embodiment of the present invention;

[0058] Figure 4 This is a training diagram of a deep learning data model in a maximum power point tracking control method for a photovoltaic power generation system according to an embodiment of the present invention;

[0059] Figure 5 This is a flowchart of maximum power point prediction and model correction in a maximum power point tracking control method for a photovoltaic power generation system according to an embodiment of the present invention;

[0060] Figure 6 This is a maximum power point tracking curve diagram in a maximum power point tracking control method for a photovoltaic power generation system according to an embodiment of the present invention. Detailed Implementation

[0061] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0062] According to an embodiment of the present invention, a maximum power point tracking control method for a photovoltaic power generation system is provided.

[0063] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a maximum power point tracking control method for a photovoltaic power generation system includes:

[0064] S1. Collect the operating data of the photovoltaic power generation system and preprocess the operating data to obtain feature data in order to construct the maximum power point prediction model.

[0065] In this optional embodiment, collecting operational data from the photovoltaic power generation system and preprocessing the operational data to obtain feature data to construct a maximum power point prediction model includes:

[0066] S11. Set up a data acquisition node between the photovoltaic array and the converter to collect the operating data of the photovoltaic power generation system in real time through a sensor network.

[0067] S12. Record the running data according to the preset sampling period and establish a time series dataset.

[0068] S13. Use wavelet transform to perform multi-scale analysis on the established time series dataset, and normalize the analysis results to obtain feature data.

[0069] It should be further noted that the collected operational data of the photovoltaic power generation system, after preprocessing, yields characteristic data including:

[0070] 1. Set up a data acquisition node between the photovoltaic array and the DC-DC converter, and collect the output voltage V, current I, irradiance S and ambient temperature T of the photovoltaic module in real time through voltage sensors, current sensors, ambient irradiance sensors and temperature sensors.

[0071] 2. The collected data is recorded and organized into a time-series dataset with a fixed sampling period (e.g., 100ms or 1s). Each sample contains input features within multiple time steps, which are used for dynamic modeling in subsequent deep learning models. Simultaneously, the collected raw data undergoes wavelet transform to decompose the signal into multiple frequency bands, especially high-frequency details. Multi-scale analysis of the signal allows for better capture of time-varying features, particularly short-term changes with higher frequencies. The Discrete Wavelet Transform (DWT) of the original signal x(t) typically yields components at multiple different scales.

[0072]

[0073] In the formula, a1(t) is the approximation coefficient for the low-frequency part, representing the stationary part of the signal; d1(t), d2(t), ... d n (t) represents the high-frequency components at different scales, indicating the detailed parts of the signal; through wavelet transform, the signal is decomposed into low-frequency approximation and high-frequency detailed parts, thus enabling the extraction of richer signal features.

[0074] 3. To improve model training quality and inference robustness, the collected raw data is normalized. After wavelet transform, normalization is performed separately for each frequency band to ensure that the feature information of each frequency band is at the same scale, thus avoiding interference from certain frequency band information to the overall signal during model training. For each frequency band a obtained from the wavelet transform... k (t) and d k (t) (where k = 1, 2, ..., n) uses the min-max normalization method to map all variables to a uniform numerical range:

[0075]

[0076] Among them, a' k (t) and d' k (t) represent the low-frequency and high-frequency components after normalization; min(a) k (t)) and max(a k (t) represents the frequency band a k The minimum and maximum values ​​of (t), similarly, min(d) k (t)) and max(d k (t)) is the detail part d k The minimum and maximum values ​​of (t) are determined; the goal of normalization is to map the values ​​of each frequency band to the range [0,1], thereby eliminating the numerical differences between different frequency bands; the normalized information of each frequency band can be directly used in the subsequent modeling stage. Based on this, the normalization results of each frequency band are recombined to obtain a multi-frequency band feature vector, which is used as the input of the subsequent deep learning model; outliers caused by sensor failures are identified and removed based on the sliding window averaging and median filtering methods; for missing data caused by communication delays or sensor failures, linear interpolation is used to fill in the gaps.

[0077] 4. During the offline training phase, the maximum power value P obtained during the acquisition period is used as the basis for the training. m Calculate its corresponding voltage V m As the target label for supervised learning, it forms an input-output data pair.

[0078] S14. Based on the obtained feature data, construct a deep learning-based maximum power point prediction model.

[0079] In this optional embodiment, constructing a deep learning-based maximum power point prediction model based on the obtained feature data includes:

[0080] S141. Based on the temporal and nonlinear characteristics of the feature data, establish a deep learning framework based on long short-term memory networks.

[0081] S142. Embed a spatial and temporal dual-path attention mechanism in the input layer of the deep learning framework to perform weighted processing on feature data in order to capture the spatial dependencies between sensors and the evolution trend of time series.

[0082] In this optional embodiment, embedding a spatial and temporal dual-path attention mechanism into the input layer of the deep learning framework to weight the feature data and capture the spatial dependencies between sensors and the evolution trend of the time series includes:

[0083] S1421. Based on each time step of the feature data, the spatial attention mechanism is used to calculate the spatial attention weight of the feature data of each sensor in order to capture the spatial dependency between different sensors at the same time.

[0084] S1422. Based on the time series of feature data, the feature data at each time node are weighted using a time attention mechanism to obtain the time attention weight, so as to capture the evolution trend of the time series.

[0085] S1423. Combine spatial attention weights and temporal attention weights to perform joint weighting on the feature data and obtain the weighted feature data.

[0086] S143. Using the weighted feature data as input and the operating point corresponding to the maximum power point as output, construct a maximum power point prediction model based on deep learning.

[0087] It should be added that, such as Figure 4 As shown, a deep learning-based maximum power point prediction model is used to quickly and accurately predict the optimal output power or voltage under different environmental conditions. The construction of the deep learning-based maximum power point prediction model includes:

[0088] 1. Based on the temporal and nonlinear mapping characteristics of photovoltaic system operation data, a Long Short-Term Memory (LSTM) network is selected as the deep learning network structure for modeling. This deep learning network structure is mainly used to extract short-term fluctuations and long-term trends in the time series. The LSTM unit structure is as follows: Figure 3 As shown.

[0089] Before the model's input layer, a spatial-temporal dual-path attention mechanism is employed to assign different weights to different types of input data, enabling the model to automatically focus on the features most relevant to the prediction of the maximum power point. Furthermore, the dual-path attention mechanism considers both spatial dependence (the relationship between different sensor data) and temporal dependence (the trend of time-series data over time), dynamically adjusting the weights of different sensor features based on varying environmental conditions and the relative relationships between sensor data. Spatial attention weight calculation:

[0090] For the sensor data X(t) = [V(t), I(t), S(t), T(t)] at time t, where V(t) is the voltage data collected by the sensor at time t, I(t) is the current data collected by the sensor at time t, S(t) is the power data collected by the sensor at time t, and T(t) is the temperature data collected by the sensor at time t, the weight A of each sensor feature is calculated using an attention network. S (t):

[0091] A s (t)=softmax(W s ·X(t));

[0092] In the formula, X(t) represents the data collected by the sensor, and W s represents the learned weight matrix; softmax represents the standard softmax function, used to map the weights to the range [0,1], ensuring that the sum of the weights of all features is 1; A S (t) represents the spatial attention weight of the data features.

[0093] Through this spatial attention mechanism, the model can dynamically adjust the weighting of input features based on the importance of each sensor data point at each time step.

[0094] The time attention mechanism focuses on how to weight different historical moments based on the temporal dependencies of time series data. The output power of a photovoltaic system is affected not only by the current environmental state but also by the historical states of past moments. By calculating the data weights for each historical moment, the time attention mechanism enables the model to focus on the historical data most useful for current predictions. For historical state data H(t) = [P] m (t-1),P m (t-2),…,P m (tk)],P m H(t) represents the output power of the photovoltaic system at time t, and H(t) is the historical power data sequence at time t, containing power data from k times. The weight A for each historical time is calculated using a time attention network. t (t):

[0095] A t (t)=softmax(W t ·H(t));

[0096] In the formula, W t This represents the learned temporal attention weight matrix; softmax is the standard softmax function; A t (t) represents the time attention weight.

[0097] 2. Define the input and output of the deep learning model. The input consists of preprocessed feature variables, including: the output voltage V, output current I, irradiance S, temperature T of the photovoltaic cell weighted by a spatial-temporal dual-path attention mechanism, their historical time series segments, and the historical power data sequence weighted by temporal attention. These calculated weights are used to adjust the weighted features of each sensor data feature. The model identifies these weighted features as input and uses their temporal characteristics to model the relationship between the output power of the photovoltaic system and related features. The output of the deep learning model is the voltage value V corresponding to the target maximum power point. m and power value P m The combination of .

[0098] 3. Based on the generated historical data samples, construct the training set and validation set required for supervised learning, and perform data augmentation or balanced sampling according to season, weather type, etc., to improve the model's generalization ability to different operating conditions.

[0099] 4. Using mean squared error (MSE) as the loss function and combining it with the Adam optimization algorithm, the neural network is iteratively trained. An early stopping strategy is introduced during the training process to prevent overfitting.

[0100] 5. The trained deep learning model is lightweighted and deployed to the edge computing module of the photovoltaic controller (such as an ARM chip, embedded system, or FPGA platform) to achieve real-time prediction. The prediction cycle is synchronized with data sampling, generally less than 1 second, meeting the fast response requirements of MPPT.

[0101] S2. The maximum power point prediction model is trained based on the acquired historical operating data samples, and the trained maximum power point prediction model is used to predict the target operating point.

[0102] In this optional embodiment, training the maximum power point prediction model based on acquired historical operating data samples, and using the trained maximum power point prediction model to predict the target operating point includes:

[0103] S21. Based on the physical characteristics of photovoltaic cells and environmental factors, establish a mechanism model for maximum power point tracking of photovoltaic power generation system to characterize the nonlinear relationship between output power and voltage.

[0104] In this optional embodiment, the expression for the mechanistic model is:

[0105]

[0106] In the formula, I represents the output current of the photovoltaic cell; PHI0 represents the photocurrent (proportional to light intensity and battery area); I0 represents the diode reverse saturation current; q represents the electron charge (approximately 1.6 × 102). -19 C); V represents the photovoltaic cell terminal voltage; R S The diode represents the series resistance; A represents the diode ideality factor; K represents the Boltzmann constant (approximately 1.38 × 10⁻⁶). -23 J / K); T represents absolute temperature (unit: K).

[0107] It should be further explained that a maximum power point tracking (MPPT) mechanism model for a photovoltaic (PV) power generation system was established. This model includes a photovoltaic current source, diodes, and series and parallel resistors, accurately describing the current-voltage (IV) characteristics of the PV cell under different temperature and illumination conditions, providing fundamental data for the subsequent MPPT algorithm. Based on the mechanism model, it can be found that under given environmental conditions, the output power P of the PV cell exhibits a nonlinear relationship with the operating voltage V, such as... Figure 2 As shown; at voltage V m At this point, the output power reaches its maximum value P. m Therefore, this point is the maximum power operating point, and the process of photovoltaic power generation maximum power point tracking is the process of adjusting the DC side operating voltage of the photovoltaic cell.

[0108] S22. Based on the established mechanism model, obtain historical operation data samples of the photovoltaic power generation system, and establish training and validation sets.

[0109] S23. Select the mean squared error as the loss function, and train the maximum power point prediction model by combining the established training set and validation set.

[0110] S24. Input the weighted feature data into the trained maximum power point prediction model to predict the target operating point corresponding to the maximum power point.

[0111] S3. Combining the target operating point with the ant colony optimization algorithm, determine the optimal operating point, and obtain the control parameters for maximum power point tracking by constructing an adaptive voltage controller.

[0112] In this optional embodiment, the optimal operating point is determined by combining the target operating point with the ant colony optimization algorithm, and the control parameters for maximum power point tracking are obtained by constructing an adaptive voltage controller, including:

[0113] S31. Initialize the individual positions and parameters in the ant colony algorithm, construct the objective function, and use the target working point as the local search center to guide the ant colony to search.

[0114] In this optional embodiment, initializing the individual positions and parameters in the ant colony algorithm, constructing the objective function, and using the target working point as the local search center to guide the ant colony in the search includes:

[0115] S311. Initialize the individual positions and parameters in the ant colony algorithm, and perturb the initial parameters using Gaussian noise;

[0116] S312. Based on the parameter distribution after perturbation, the difference between the predicted operating point and the actual operating point of the photovoltaic power generation system is used as the optimization objective to construct an objective function;

[0117] S313. Use the predicted working point as the local search center to guide the ant colony to perform a local search within the search space, and combine evolutionary strategies to optimize the search process.

[0118] S32. Based on the search results from the ant colony, dynamically optimize the parameters and update the pheromone concentration of the search path according to the deviation between the target working point and the actual working point.

[0119] S33. Based on the pheromone concentration update results, the search path of individual ants is corrected through a feedback mechanism to obtain the global optimal solution and the optimal working point.

[0120] It should be added that, such as Figure 4 and Figure 5 As shown, combining the target working point and the ant colony optimization algorithm, the optimal working point is determined as follows:

[0121] 1. Initialize the ant colony optimization algorithm:

[0122] (1) Initialize the ant colony: There are N ants, and the initial position of each ant in the search space is determined by the voltage value V. i and its corresponding power change ΔP i Decide:

[0123] X i =(V i ,ΔP i );

[0124] In the formula, V i Let ΔP represent the voltage value of the i-th ant. i This indicates the change in its power.

[0125] To ensure that each ant possesses a specific genotype, the genes of each ant are defined as a set of parameters, including search step size, pheromone update factor, etc. The ant's genes are represented as follows:

[0126]

[0127] In the formula, V i ΔV represents the voltage value of the i-th ant. i Indicates its search step size; This represents the pheromone update factor.

[0128] For the genes of the initial individual, Gaussian noise is used to perturb each gene to maintain a certain degree of diversity:

[0129]

[0130] Where N(0,σ) 2 ) represents the standard normally distributed noise, and σ is the standard deviation of the disturbance.

[0131] (2) Setting the objective function: The objective function is the process of optimizing control parameters. The ant colony optimization algorithm is used to find the voltage value V that maximizes the power of the photovoltaic system. m The objective function f(X) is defined as:

[0132] f(X) = P max -P current ;

[0133] In the formula, Pmax is the maximum power point power obtained based on the prediction results, P current It is the current system output power.

[0134] (3) Search space selection and evolutionary operation: Based on the predictions of the deep learning model, ant colony optimization will start from the predicted maximum power point voltage V. m Departure, around V m A local search is performed; the evolutionary strategy adjusts and optimizes the ant's path to avoid getting trapped in local optima. During the search, evolutionary operations include selection, crossover, and mutation.

[0135] A. Selection: Select high-performing ants based on the fitness function and eliminate low-performing ants.

[0136] B. By merging the genes of multiple superior individuals through crossover, new individuals are generated; the crossover operation can be represented as: Where μ is the crossover factor, ω i and ω j It's the genes of two ants.

[0137] C. Mutation: Making slight mutations to individuals to generate new search directions; Similar to the previous perturbation formula, it is used to explore the current solution space.

[0138] 2. Ant colony optimization: dynamically adjust control parameters.

[0139] (1) Step size (perturbation amount) adjustment:

[0140] The step size ΔV is a key parameter affecting the convergence speed of MPPT. In conventional methods, the step size is fixed, while in this invention, the step size is dynamically adjusted based on the deep learning prediction results. In cases of sudden changes in illumination or significant changes in system state, an evolutionary operation is used to achieve dynamic adjustment. The step size ΔV is generated by the evolutionary strategy.

[0141] ΔV i =ΔV i-1 +β·N(0,σ 2 );

[0142] Here, β is a factor that controls the step size adjustment and is adjusted based on the ant's fitness performance; this step size adjustment method ensures that the system can respond quickly when predicting drastic changes in light intensity, and make fine adjustments when the light intensity changes are stable.

[0143] (2) Step size optimization using incremental conductivity method:

[0144] The incremental conductance method determines whether the maximum power point has been reached by comparing the rates of change of voltage and current. In traditional methods, the step size is fixed. However, in this invention, the incremental conductance method uses a smaller step size as the system approaches the maximum power point to avoid excessive disturbance. The step size ΔV INC Dynamically adjusted using ant colony optimization algorithm:

[0145]

[0146] Where β is a factor used to adjust the step size in ant colony optimization. It is the derivative of the current voltage and power. Using the prediction results from deep learning, the step size can be adjusted in advance, avoiding the situation where the step size adjustment is too slow or too fast in traditional methods.

[0147] (3) Adaptive adjustment of convergence threshold:

[0148] The convergence threshold ε is defined as the threshold at which the maximum power point is considered to have been reached when the system error is less than this threshold. The convergence threshold is also dynamically adjusted based on the predicted trend of illumination change.

[0149] ε=δ·|P m -P current |;

[0150] Where ε is an adaptive factor, P m It is the predicted maximum power point power, P current This is the current power; by adjusting the convergence threshold, the system can converge faster when there are large fluctuations in illumination, while maintaining a precise maximum power point in a stable state.

[0151] (3) Pheromone updating and feedback correction:

[0152] In the ant colony optimization algorithm, the pheromone concentration τ determines the search priority; ants tend to choose paths with higher pheromone concentrations for searching. The initial pheromone concentration is: Subsequently, the pheromones will be updated based on the ants' search results: τ i+1 =(1-ρ)·τ i +Δτ; where ρ is the pheromone evaporation factor, set at 0.1≤ρ≤0.9, Δτ is the amount of pheromone left by the ant on the path, and in the optimization of the evolutionary strategy, the adjustment of pheromone is affected by the fitness of the ant; ants with higher fitness will accumulate more pheromone during the search process, guiding other ants to move closer to the optimal solution, and N is the number of ants in the colony.

[0153] Based on the prediction error of the deep learning model, the system adjusts the pheromone update mechanism in real time. If the prediction error is large (greater than 5%), the pheromone update amplitude will be increased, thereby enhancing the ants' exploratory nature during the search. If the prediction error is small (less than 1%), the pheromone update amplitude will be decreased, thereby enhancing the ants' utilization. Ant colony optimization corrects the search path based on actual feedback, ensuring that the system always stays near an optimal voltage and power point. Through this correction, local optima caused by improper parameter adjustment in traditional methods can be avoided.

[0154] When the power change of the system is less than the convergence threshold (set to 10⁻³), it is considered that the system has found the maximum power point and entered a steady state. Under stable or slowly changing illumination conditions, the system will remain at the optimal maximum power point. When the illumination changes abruptly, deep learning prediction and ant colony optimization will quickly adjust the control parameters and re-track the optimal power point. Finally, through repeated iterations of the adjustment mechanism of ant colony optimization and deep learning feedback, the operating voltage and power near the maximum power point are obtained.

[0155] S34. Based on the optimal operating point, construct an adaptive voltage controller to obtain the control parameters for maximum power point tracking.

[0156] In this optional embodiment, an adaptive voltage controller is constructed based on the optimal operating point to obtain the control parameters for maximum power point tracking, including:

[0157] S341. Based on the obtained optimal operating point, design an adaptive voltage controller that integrates deep reinforcement learning and meta-learning frameworks to generate duty cycle control signals for driving the converter.

[0158] In this optional embodiment, based on the obtained optimal operating point, an adaptive voltage controller integrating deep reinforcement learning and meta-learning frameworks is designed to generate duty cycle control signals for driving the converter, including:

[0159] S3411. Construct a reinforcement learning control strategy based on a deep Q-network framework, combine the obtained optimal operating point, design a reward function with power increment as the objective, and iteratively update the reinforcement learning control strategy through the Q-learning algorithm.

[0160] S3412. Based on the meta-learning mechanism, using the model-independent meta-learning framework, the initial parameters of the strategy suitable for different environmental changes are learned through multi-task training in various photovoltaic power generation scenarios.

[0161] S3413. Integrate the initial parameters of the strategy into the updated reinforcement learning control strategy to generate a control strategy, and output the duty cycle control signal for driving the converter based on the control strategy.

[0162] S342. The generated duty cycle control signal is applied to the converter in real time to dynamically adjust the converter's duty cycle to control the output voltage. The control process is optimized by combining the jitter suppression mechanism and the anti-integral saturation mechanism to obtain the control parameters for maximum power point tracking.

[0163] It should be added that, such as Figure 6 As shown, based on the optimal operating point, an adaptive voltage controller is constructed to obtain the control parameters for maximum power point tracking, including:

[0164] 1. Definition of state space and action space:

[0165] (1) State space (S): includes the current environmental parameters of the system (irradiance, temperature) and the current voltage and power values ​​of the photovoltaic system: S={V t ,P t G t ,T t}, where V t It is the current voltage, P t It is the current power, G t This is the current irradiance, T t This is the current temperature.

[0166] (2) Action Space (A): The voltage regulation action that the controller can select, which can be to increase or decrease the voltage amplitude: A = {ΔV} t}, where ΔV t This indicates that control actions can be learned and adjusted through deep reinforcement learning algorithms.

[0167] 2. Design of a control strategy that integrates deep reinforcement learning and meta-learning:

[0168] (1) Design of deep reinforcement learning control strategy: In this controller, a deep Q network is used to train the agent. The goal of the agent is to optimize the voltage regulation strategy Q(S,A) so that the power output of the photovoltaic system is as close as possible to the maximum power point.

[0169] A. Reward Function Design: The reward function is designed based on changes in system power, with the goal of increasing power output through voltage regulation and stabilizing it near the maximum power point. The reward function can be expressed as:

[0170] R t =ΔP t =P t+1 -P t ;

[0171] Among them, P t It is the current power, ΔP t It represents the power increment; the reward design aims to encourage the agent to increase power, so if the current adjustment action increases power, the reward is positive.

[0172] B. Deep Q-Network Update: The control policy is updated using a Q-learning update formula;

[0173]

[0174] Where η is the learning rate, λ is the discount factor, and A' is the next action.

[0175] (2) Meta-learning task design:

[0176] Deep reinforcement learning enables agents to learn how to dynamically adjust voltage based on the current state through environmental feedback, thereby stabilizing it near the maximum power point. Meta-learning aims to enable agents to fine-tune existing strategies under new environmental conditions. Therefore, it relies on the initially trained control strategy and uses a small number of samples for rapid fine-tuning without retraining the entire control model. Meta-learning helps the controller quickly update parameters when the environment changes through its rapid adaptability. Specifically, the controller can use meta-learning to adjust its control strategy with a small number of training samples to cope with sudden environmental changes. By simulating different environmental changes (changes in light intensity and temperature), model-independent meta-learning methods are used during training to learn adaptive adjustment strategies.

[0177]

[0178] Among them, L(E i ,θ) represents different environmental conditions E i The loss function is given by θ, where θ represents the voltage regulation error. * The optimized meta-model parameters can be quickly fine-tuned in new tasks to cope with rapidly changing environments.

[0179] The meta-learning component fine-tunes the deep reinforcement learning strategy, enabling the controller to quickly adapt to different environments. The loss in this process is constrained and updated using a Q-value function. Meta-learning minimizes the loss function, ensuring the control strategy operates stably and efficiently across multiple environments. Deep reinforcement learning, on the other hand, optimizes the Q-value of voltage regulation through interaction with the environment. Deep reinforcement learning acts as the main driver of the controller, dynamically adjusting the voltage based on the current state and historical feedback of the photovoltaic system, optimizing the system's power output in real time. The agent is trained by maximizing the reward function, enabling it to adjust the voltage according to different environmental changes, approximating the maximum power point. Meta-learning is responsible for rapidly adapting to environmental changes. In practical applications, the environment of photovoltaic systems fluctuates significantly. Meta-learning learns how to quickly adjust the strategy, allowing the controller to adapt to new operating conditions with minimal training or samples, avoiding the need for deep reinforcement learning to be retrained from scratch.

[0180] (3) DC-DC converter drive: The controller output signal D(t) is ultimately used to drive the DC-DC converter in the photovoltaic system. By adjusting the duty cycle, the output voltage V is controlled to achieve rapid approximation of V. m .

[0181] (4) To prevent high-frequency jitter during the control process, a jitter suppression mechanism is set: a dead zone is set so that the duty cycle is not updated at that time; an anti-integral saturation mechanism is introduced to limit the growth of the integral term. Through the controller design, the system can not only respond quickly to the prediction results of the deep learning + mechanism model, but also has the ability to adaptively adjust to abnormal disturbances, thereby improving the energy efficiency and stability of the photovoltaic MPPT system in dynamic environments.

[0182] In a specific embodiment, a maximum power point tracking control method for a photovoltaic power generation system is as follows:

[0183] A small-scale photovoltaic power generation system is adopted, and its key components include:

[0184] (1) Photovoltaic module: nominal power 300W, open circuit voltage 42.4V, short circuit current 8.8A; (2) DC-DC converter: adopts Boost structure, input voltage range 15~40V, maximum output power 400W; (3) Controller: based on STM32F407 chip, main frequency 168MHz, equipped with AI model inference engine (using CMSIS-NN and ONNX-Runtime); (4) Sensor: voltage and current sensor sampling frequency 1kHz; external temperature and irradiance sensors; (5) host computer: used for data recording, model training and deployment control.

[0185] 2. The deep learning model adopts a two-layer LSTM + fully connected layer network structure; input variables: voltage, current, irradiance, temperature and their historical data (length 10); output variable: the voltage at the maximum power point V at the next time step. m The training platform used was PyTorch, with 200,000 sets of data. The model was deployed and quantized to float16, then converted to an ONNX model and embedded into an STM32 for execution. Initial K... p =0.8,K i =0.01; if the error does not converge within 3 periods, K will be automatically adjusted. p Up to 0.6; jitter suppression dead zone ε = 0.5V, no adjustment is made if the error is below this value; operating condition settings: initial irradiance: 600W / m 2 The voltage suddenly jumped to 1000 W / m at 1.5 seconds. 2 The system records the power output process for 4 seconds.

[0186] The maximum power point tracking curve of a photovoltaic cell is as follows: Figure 5 As shown; response time (mutation recovery) 0.8s; steady-state power tracking error ≤1.5%; average power output improvement +8.7%; jitter amplitude (standard deviation) is ±1.1W.

[0187] Furthermore, deep learning prediction models are used for maximum power point (MPPT) identification in photovoltaic (PV) power generation. A deep learning model (LSTM) is introduced into PV MPPT control to predict the voltage or power at the MPPT, enabling rapid perception and dynamic prediction of the MPPT under nonlinear and multi-disturbance environments, significantly improving response speed and prediction accuracy. Mechanistic models assist in correcting AI prediction outputs. A PV cell mechanism model is introduced on top of AI deep learning predictions to construct a multi-source information fusion framework, enabling prediction correction based on physical laws. This improves system stability and interpretability, mitigating the failure risk of purely data-driven models in extreme scenarios. The definition and use of the prediction confidence index γ, based on real-time prediction error, effectively controls the AI ​​model's dominance by dynamically adjusting the weight of deep learning predictions in the fusion process. This ensures the system reverts to physical model dominance when confidence is low, improving control safety and robustness. The integrated design of prediction and dynamic control uses the fusion prediction result as the control target input and introduces a gain-adjustable PI controller or fuzzy controller to achieve adaptive voltage regulation control. The controller can automatically adjust parameters based on error trends, prediction confidence and other indicators to improve tracking accuracy and robustness.

[0188] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for maximum power point tracking control in a photovoltaic power generation system, characterized in that, The method includes: S1. Collect the operating data of the photovoltaic power generation system and preprocess the operating data to obtain feature data in order to construct the maximum power point prediction model; S2. Train the maximum power point prediction model based on the acquired historical operating data samples, and use the trained maximum power point prediction model to predict the target operating point. S3. Combining the target operating point with the ant colony optimization algorithm, determine the optimal operating point, and obtain the control parameters for maximum power point tracking by constructing an adaptive voltage controller.

2. The maximum power point tracking control method for a photovoltaic power generation system according to claim 1, characterized in that, The process of collecting operational data from the photovoltaic power generation system and preprocessing the operational data to obtain feature data for constructing a maximum power point prediction model includes: S11. Set up a data acquisition node between the photovoltaic array and the converter, and collect the operating data of the photovoltaic power generation system in real time through a sensor network; S12. Record the running data according to the preset sampling period and establish a time series dataset; S13. Apply wavelet transform to perform multi-scale analysis on the established time series dataset, and normalize the analysis results to obtain feature data. S14. Based on the obtained feature data, construct a deep learning-based maximum power point prediction model.

3. The maximum power point tracking control method for a photovoltaic power generation system according to claim 2, characterized in that, The step of constructing a deep learning-based maximum power point prediction model based on the obtained feature data includes: S141. Based on the temporal and nonlinear characteristics of the feature data, establish a deep learning framework based on long short-term memory networks. S142. Embed a spatial and temporal dual-path attention mechanism in the input layer of the deep learning framework to weight the feature data in order to capture the spatial dependencies between sensors and the evolution trend of the time series. S143. Using the weighted feature data as input and the operating point corresponding to the maximum power point as output, construct a maximum power point prediction model based on deep learning.

4. The maximum power point tracking control method for a photovoltaic power generation system according to claim 3, characterized in that, The method of embedding a spatial and temporal dual-path attention mechanism in the input layer of the deep learning framework to weight feature data and capture the spatial dependencies between sensors and the evolution trend of time series includes: S1421. Based on each time step of the feature data, the spatial attention mechanism is used to calculate the spatial attention weight of the feature data of each sensor in order to capture the spatial dependency between different sensors at the same time. S1422. Based on the time series of feature data, the feature data at each time node are weighted using a time attention mechanism to obtain the time attention weight, so as to capture the evolution trend of the time series. S1423. Combine spatial attention weights and temporal attention weights to perform joint weighting on the feature data and obtain the weighted feature data.

5. The maximum power point tracking control method for a photovoltaic power generation system according to claim 1, characterized in that, The process of training the maximum power point prediction model based on acquired historical operating data samples, and then using the trained maximum power point prediction model to predict the target operating point, includes: S21. Based on the physical characteristics of photovoltaic cells and environmental factors, establish a mechanism model for maximum power point tracking of photovoltaic power generation system to characterize the nonlinear relationship between output power and voltage. S22. Based on the established mechanism model, obtain historical operation data samples of the photovoltaic power generation system, and establish training and validation sets; S23. Select the mean square error as the loss function, and train the maximum power point prediction model by combining the established training set and validation set. S24. Input the weighted feature data into the trained maximum power point prediction model to predict the target operating point corresponding to the maximum power point.

6. The maximum power point tracking control method for a photovoltaic power generation system according to claim 5, characterized in that, The expression for the mechanism model is: In the formula, I represents the output current of the photovoltaic cell; PH Represents the photogenerated current; I0 ​​represents the diode reverse saturation current; q represents the electron charge; V represents the photovoltaic cell terminal voltage; R S A represents the series resistance; A represents the diode ideality factor; K represents the Boltzmann constant; T represents the absolute temperature.

7. The maximum power point tracking control method for a photovoltaic power generation system according to claim 1, characterized in that, The process of combining the target operating point with the ant colony optimization algorithm to determine the optimal operating point and obtaining the control parameters for maximum power point tracking by constructing an adaptive voltage controller includes: S31. Initialize the individual positions and parameters in the ant colony algorithm, construct the objective function, and use the target working point as the local search center to guide the ant colony to search; S32. Based on the search results from the ant colony, dynamically optimize the parameters and update the pheromone concentration of the search path according to the deviation between the target working point and the actual working point. S33. Based on the pheromone concentration update results, the search path of individual ants is corrected through a feedback mechanism to obtain the global optimal solution and the optimal working point. S34. Based on the optimal operating point, construct an adaptive voltage controller to obtain the control parameters for maximum power point tracking.

8. The maximum power point tracking control method for a photovoltaic power generation system according to claim 7, characterized in that, The process of initializing the individual positions and parameters in the ant colony algorithm, constructing the objective function, and using the target working point as the local search center to guide the ant colony in the search includes: S311. Initialize the individual positions and parameters in the ant colony algorithm, and perturb the initial parameters using Gaussian noise; S312. Based on the parameter distribution after perturbation, the difference between the predicted operating point and the actual operating point of the photovoltaic power generation system is used as the optimization objective to construct an objective function; S313. Use the predicted working point as the local search center to guide the ant colony to perform a local search within the search space, and combine evolutionary strategies to optimize the search process.

9. The maximum power point tracking control method for a photovoltaic power generation system according to claim 8, characterized in that, The process of constructing an adaptive voltage controller based on the optimal operating point to obtain the control parameters for maximum power point tracking includes: S341. Based on the obtained optimal operating point, design an adaptive voltage controller that integrates deep reinforcement learning and meta-learning frameworks to generate duty cycle control signals for driving the converter. S342. The generated duty cycle control signal is applied to the converter in real time to dynamically adjust the converter's duty cycle to control the output voltage. The control process is optimized by combining the jitter suppression mechanism and the anti-integral saturation mechanism to obtain the control parameters for maximum power point tracking.

10. The maximum power point tracking control method for a photovoltaic power generation system according to claim 9, characterized in that, Based on the obtained optimal operating point, an adaptive voltage controller integrating deep reinforcement learning and meta-learning frameworks is designed to generate a duty cycle control signal for driving the converter, including: S3411. Construct a reinforcement learning control strategy based on a deep Q-network framework, combine the obtained optimal operating point, design a reward function with power increment as the objective, and iteratively update the reinforcement learning control strategy through the Q-learning algorithm. S3412. Based on the meta-learning mechanism, using the model-independent meta-learning framework, the initial parameters of the strategy suitable for different environmental changes are learned through multi-task training in various photovoltaic power generation scenarios. S3413. Integrate the initial parameters of the strategy into the updated reinforcement learning control strategy to generate a control strategy, and output the duty cycle control signal for driving the converter based on the control strategy.

Images