Artificial intelligence-based solar panel light tracking optimization method and system

Abstract

The application provides a solar panel light tracking optimization method and system based on artificial intelligence, relates to the technical field of solar energy utilization, and comprises the following steps: collecting geographic position, time information and light intensity data; extracting features through empirical mode decomposition to generate a layered light prediction sequence; calculating the optimal tracking angle by using differential evolution operation; monitoring the surface temperature in real time, generating a dynamic obstacle avoidance scheme based on temperature gradient field analysis; and optimizing the tracking strategy through continuous power feedback. The application improves the power generation efficiency of the solar panel under complex light conditions and enhances the adaptability of the system to environmental changes.

Description

Technical Field

[0001] This invention relates to the field of solar energy utilization technology, and in particular to a method and system for optimizing solar panel illumination tracking based on artificial intelligence. Background Technology

[0002] With the advancement of global energy structure adjustments and clean energy development strategies, solar energy, as a renewable and clean energy source, has received widespread attention. Solar panels, as the core equipment for solar energy utilization, directly impact energy conversion efficiency and power generation through their solar tracking efficiency. Traditional solar tracking systems primarily rely on fixed-angle installations or simple timed adjustment mechanisms, which cannot adapt to complex and ever-changing lighting environments, resulting in low energy utilization efficiency. With the rapid development of artificial intelligence technology, applying intelligent algorithms to solar panel solar tracking systems has become an important direction for improving solar energy utilization efficiency.

[0003] Traditional tracking systems often employ fixed models or simple photosensitive element feedback control, lacking deep learning and prediction capabilities for light variation patterns. This prevents them from proactively adjusting to changes in light intensity under complex weather conditions, resulting in low tracking efficiency in cloudy or rapidly changing lighting environments. Existing tracking algorithms often employ single-target optimization designs, primarily considering instantaneous light intensity while ignoring comprehensive factors such as temperature distribution and the influence of obstructions, making it difficult to achieve globally optimal tracking performance in complex environments. Furthermore, the lack of intelligent dynamic obstacle avoidance mechanisms means that when solar panels are temporarily obstructed or subjected to localized shadows, they cannot quickly identify and adjust their posture, easily causing hot spot effects and a sharp decline in power generation efficiency, potentially even leading to equipment damage and affecting the long-term stable operation of the system.

[0004] In summary, there is an urgent need for an AI-based method for optimizing solar panel illumination tracking. This method should utilize multi-sensor data fusion, intelligent prediction algorithms, and adaptive control strategies to achieve efficient tracking and dynamic adjustment of solar panels, thereby improving solar energy utilization efficiency. Summary of the Invention

[0005] This invention provides an artificial intelligence-based method and system for optimizing solar panel illumination tracking, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides an artificial intelligence-based method for optimizing solar panel illumination tracking, comprising:

[0007] Collect geographical location and time information of the solar panel installation location, and obtain light intensity data and current output power data of the solar panel through multiple light sensors;

[0008] Empirical mode decomposition is performed on the illumination intensity data to obtain intrinsic components. Features are extracted based on the frequency characteristics of the intrinsic components to generate a hierarchical illumination prediction sequence.

[0009] Differential evolution operation is performed using historical output power data and layered illumination prediction sequence, combined with adaptive adjustment of mutation factor, to calculate the optimal tracking angle group;

[0010] The optimal tracking angle group is input into the dual-axis drive device to perform solar panel attitude adjustment and collect surface temperature distribution in real time. When an abnormal temperature is detected, the temperature gradient field analysis is used to calculate the movement characteristics of the obstruction and combine it with the layered illumination prediction sequence to generate a dynamic obstacle avoidance scheme.

[0011] The solar panel attitude is adjusted according to the dynamic obstacle avoidance scheme, and the adjusted output power data is collected as input data for the next round of differential evolution calculation.

[0012] In one optional embodiment, performing empirical mode decomposition on the illumination intensity data to obtain intrinsic components, extracting features based on the frequency characteristics of the intrinsic components, and generating a hierarchical illumination prediction sequence includes:

[0013] The light intensity data is segmented, and the maximum and minimum points of each segment are calculated. The upper and lower envelopes are generated by cubic spline interpolation.

[0014] Calculate the mean values ​​of the upper envelope and the lower envelope, and subtract the mean values ​​from the illumination intensity data to obtain the initial residual signal;

[0015] Repeatedly extract time-frequency features and calculate decomposition quality scores from the initial residual signal, and obtain new residual signals through envelope extraction and mean calculation until the intrinsic components are obtained;

[0016] The inherent components are sorted by frequency. The inherent component with the lowest frequency is used as the trend component to calculate the rate of change of the signal to obtain the change characteristics. The abrupt change point position and depth of the inherent component that is adapted to the movement frequency of the obstruction are used as the obstruction characteristics. The fluctuation period and amplitude of the remaining high-frequency inherent components are used as the disturbance characteristics.

[0017] The change prediction value is calculated based on the change characteristics, the periodic prediction value is calculated based on the occlusion characteristics, and the fluctuation prediction value is calculated based on the disturbance characteristics. The change prediction value, the periodic prediction value, and the fluctuation prediction value are superimposed to generate a layered illumination prediction sequence.

[0018] In one optional embodiment, the time-frequency feature extraction and decomposition quality score calculation are repeatedly performed on the initial residual signal, and a new residual signal is obtained through envelope extraction and mean calculation, until the intrinsic components are obtained, including:

[0019] The initial residual signal is segmented according to a sliding window, and the time-domain statistics and frequency-domain power spectrum within each window are extracted to construct a time-frequency feature matrix. The time-frequency feature matrix is ​​then input into a pre-trained deep neural network, which identifies nonlinear feature patterns in the initial residual signal through a multi-layer attention mechanism and outputs a feature similarity distribution map.

[0020] The stationarity index, spectral entropy, and fractal dimension of the residual signal are calculated based on the feature similarity distribution map. Fuzzy membership functions are set for the stationarity index, spectral entropy, and fractal dimension. Fuzzy rule inference is performed to obtain the current decomposition quality score.

[0021] The current decomposition quality score and time-frequency feature matrix are stored in a dynamic memory pool. A reinforcement learning method is used to establish a mapping relationship between the score and the time-frequency feature matrix. Based on the mapping relationship, the quality improvement value of the next round of decomposition is calculated, and the current termination probability value is generated.

[0022] When the current termination probability value is greater than the adaptive threshold, the decomposition process is stopped and all the obtained intrinsic components are output; when the current termination probability value is less than or equal to the adaptive threshold, the fuzzy rule weights are updated, and a new residual signal is obtained through envelope extraction and mean calculation; the adaptive threshold is updated based on all historical decomposition scores stored in the memory pool.

[0023] In one optional embodiment, differential evolution is performed using historical output power data and hierarchical illumination prediction sequences, combined with adaptive adjustment of the mutation factor, to calculate the optimal tracking angle set, including:

[0024] Historical output power data and hierarchical illumination prediction sequences are used as input data for differential evolution. Each tracking angle group is set to include azimuth and tilt angles. Initial population individuals are generated by random sampling within a preset range. Each initial population individual includes an initial azimuth and an initial tilt angle.

[0025] The Euclidean distance between individuals in the initial population is calculated to obtain the inter-individual distance matrix. The mean and variance of the inter-individual distance matrix are statistically analyzed to determine the population clustering degree. The adjustment step size of the variation factor is calculated based on the population clustering degree. The adjustment step size is superimposed with the current variation factor to form an adaptive adjustment term for the variation factor.

[0026] The global optimal individual is selected from the initial population as the basis vector. Two different individuals are selected to calculate the difference vector. The mutation factor adaptive adjustment term is multiplied by the difference vector and superimposed on the basis vector to generate the mutated individual.

[0027] Calculate the predicted power values ​​of the mutated individuals and the initial population individuals respectively, select the individual corresponding to the maximum predicted power value to enter the next generation population, and repeat the process until the optimal tracking angle group is obtained.

[0028] In an optional embodiment, it further includes:

[0029] Based on the initial population individuals, the initial value and mapping equation for iterative calculation are determined. The initial value is substituted into the mapping equation to obtain the first iteration result. The first iteration result is used as input and substituted into the mapping equation again to obtain the second iteration result. The iteration is repeated until the number of iterations reaches the size of the initial population. Each iteration result is a chaotic value between zero and one.

[0030] The adaptive coefficient is determined based on the number of generations of the current population. The adaptive coefficient decreases as the number of generations increases. The adaptive coefficient is multiplied by the chaotic value to determine the chaotic perturbation term.

[0031] The chaotic perturbation term is allocated to the azimuth angle and tilt angle according to a preset ratio to perturb the initial population individuals, thereby obtaining the perturbed population individuals.

[0032] When calculating the new generation population, based on the preset number of iterations, the relationship between the increase in the optimal predicted output power value of multiple consecutive generations of population and the preset magnitude threshold is determined, and the adaptive coefficient is adjusted according to the determination result.

[0033] In one optional embodiment, when a temperature anomaly is detected, the movement characteristics of the obstruction are calculated using temperature gradient field analysis, and a dynamic obstacle avoidance scheme is generated by combining the illumination prediction sequence, including:

[0034] Collect temperature distribution data on the surface of the solar panel, calculate the mean and standard deviation of the temperature distribution data at different time points, and obtain the temperature distribution baseline value under normal operating conditions;

[0035] The temperature distribution data is subtracted from the temperature distribution benchmark value to obtain the temperature difference data, and the temperature anomaly area is determined based on the time change of the temperature difference data.

[0036] Boundary extraction is performed on the temperature anomaly region, the temperature gradient value of the temperature anomaly region is calculated, adjacent temperature gradient values ​​are clustered, and the centroid coordinates and area of ​​the clustered region are calculated to obtain the initial spatial position of the occlusion.

[0037] Calculate the temperature gradient values ​​in the horizontal and vertical directions of the temperature anomaly region to obtain the temperature gradient field, and record the gradient change values ​​of the temperature gradient field over a continuous time period.

[0038] The direction of the largest temperature change is extracted from the temperature gradient field as the main direction. Feature points are selected along the main direction, and the position changes of the feature points at different times are tracked to determine the contour change of the occluder. The moving direction and moving speed of the occluder are calculated based on the contour change.

[0039] The location coordinates of the obstruction in the future time period are calculated based on the direction and speed of movement, and the location coordinates are combined with the layered illumination prediction sequence to calculate the obstruction area.

[0040] Within a preset time window, the obstacle avoidance adjustment angle is calculated based on the occlusion area. The obstacle avoidance adjustment angle is then executed in segments. After each segment is executed, the obstacle avoidance adjustment angle for the next segment is calculated based on the new occlusion area, until the entire obstacle avoidance process is completed.

[0041] A second aspect of the present invention provides an artificial intelligence-based solar panel illumination tracking optimization system, comprising:

[0042] The first unit is used to collect geographical location information and time information of the solar panel installation location, and to obtain light intensity data and current output power data of the solar panel through multiple light sensors;

[0043] The second unit is used to perform empirical mode decomposition on the illumination intensity data to obtain intrinsic components, extract features based on the frequency characteristics of the intrinsic components, and generate a hierarchical illumination prediction sequence.

[0044] The third unit is used to perform differential evolution operations using historical output power data and hierarchical illumination prediction sequences, combined with adaptive adjustment of mutation factors, to calculate the optimal tracking angle group;

[0045] The fourth unit is used to input the optimal tracking angle set into the dual-axis drive device to perform solar panel attitude adjustment, collect surface temperature distribution in real time, and when an abnormal temperature is detected, use temperature gradient field analysis to calculate the movement characteristics of the obstruction, and combine the layered illumination prediction sequence to generate a dynamic obstacle avoidance scheme.

[0046] The fifth unit is used to adjust the solar panel attitude according to the dynamic obstacle avoidance scheme and collect the adjusted output power data as input data for the next round of differential evolution calculation.

[0047] A third aspect of the present invention provides an electronic device, comprising:

[0048] processor;

[0049] Memory used to store processor-executable instructions;

[0050] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0051] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0052] In this embodiment of the invention, by collecting geographical location, time information, and data from multiple light sensors, and combining this with empirical mode decomposition technology to extract light features, a layered light prediction sequence is generated. This provides a data foundation for the precise positioning of solar panels, improving the system's adaptability to changes in light and its prediction accuracy. Employing a differential evolution algorithm combined with a mutation factor adaptive adjustment mechanism, the optimal tracking angle can be quickly calculated, avoiding the limitations of traditional fixed-parameter optimization methods. This allows the system to achieve optimal power generation efficiency under various light conditions, reducing energy loss. By collecting surface temperature distribution in real time and performing temperature gradient field analysis, the system can intelligently identify obstructions and generate dynamic obstacle avoidance schemes. This ensures the safe operation of the solar panels while continuously optimizing output power, improving the system's reliability and adaptability. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the solar panel illumination tracking optimization method based on artificial intelligence, as described in an embodiment of the present invention.

[0054] Figure 2 Here is a flowchart of the residual signal decomposition process;

[0055] Figure 3 Flowchart for solar tracking angle optimization using differential evolution and chaotic perturbations. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0058] Figure 1 This is a flowchart illustrating the artificial intelligence-based solar panel illumination tracking optimization method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0059] Collect geographical location and time information of the solar panel installation location, and obtain light intensity data and current output power data of the solar panel through multiple light sensors;

[0060] Empirical mode decomposition is performed on the illumination intensity data to obtain intrinsic components. Features are extracted based on the frequency characteristics of the intrinsic components to generate a hierarchical illumination prediction sequence.

[0061] Differential evolution operation is performed using historical output power data and layered illumination prediction sequence, combined with adaptive adjustment of mutation factor, to calculate the optimal tracking angle group;

[0062] The optimal tracking angle group is input into the dual-axis drive device to perform solar panel attitude adjustment and collect surface temperature distribution in real time. When an abnormal temperature is detected, the temperature gradient field analysis is used to calculate the movement characteristics of the obstruction and combine it with the layered illumination prediction sequence to generate a dynamic obstacle avoidance scheme.

[0063] The solar panel attitude is adjusted according to the dynamic obstacle avoidance scheme, and the adjusted output power data is collected as input data for the next round of differential evolution calculation.

[0064] In one optional embodiment, geographic location and time information of the solar panel installation location are collected. Geographic location information includes longitude, latitude, and altitude, obtained through a built-in GPS module or manual input; time information includes the current date, time, and time zone, obtained through the system's built-in clock module. Light intensity data is acquired using multiple light sensors distributed at different locations on the solar panel. These sensors are silicon photovoltaic type light sensors capable of detecting light intensity in the 300-1100 nanometer wavelength range, with a sampling frequency of once per second. The output current and voltage of the solar panel are collected in real time using current and voltage sensors to calculate the current output power data.

[0065] Empirical Mode Decomposition (EMD) is performed on the collected illumination intensity data. This process first treats the original illumination intensity time series as a complex signal containing multiple oscillation modes, and decomposes it into multiple intrinsic mode functions (IMFs) and a residual function through repeated screening. All local extrema are identified, and upper and lower envelopes are generated using spline interpolation. The mean envelope is calculated, and the mean is subtracted from the original signal to obtain the first candidate IMF. Then, it is checked whether this candidate function meets the definition conditions of an IMF (the number of extrema is equal to or differs from the number of zero-crossings by no more than 1, and the local mean is zero across the entire data range). If not, the above steps are repeated until the conditions are met. After obtaining the first IMF, it is subtracted from the original signal, and the above process is repeated for the remaining part until the residual function becomes a monotonic function or its value is sufficiently small. Based on the obtained IMFs, their frequency characteristics are analyzed, and the average period, energy distribution, and instantaneous frequency of each IMF are calculated as characteristic parameters. Based on these characteristic parameters, illumination changes are divided into rapid changes (such as short-term cloud cover), medium-term changes (such as the movement of the sun's position during the day), and long-term changes (such as seasonal changes), generating hierarchical illumination prediction sequences with multiple time scales, ranging from short-term (within 5 minutes) to medium-term (within 24 hours).

[0066] Differential evolution (DE) is performed using historical output power data and hierarchical illumination prediction sequences to calculate the optimal tracking angle set for solar panels. The DE algorithm first initializes the population, with each individual representing a set of possible solar panel attitude angles (horizontal and pitch angles). Then, it calculates the fitness function value for each individual, which comprehensively considers energy harvesting efficiency, attitude adjustment energy consumption, and mechanical limitations under predicted illumination conditions. Next, a mutation operation is performed, generating a mutation vector for each individual through weighted differential averaging of three randomly selected individuals. Then, a crossover operation is performed, probabilistically crossing the original individuals and mutation vectors to generate trial vectors. Finally, a selection operation is performed, comparing the fitness of the trial vectors and the original individuals, retaining the better ones for the next generation. During DE, an adaptive adjustment mechanism for the mutation factor is employed, dynamically adjusting its size based on population diversity and convergence status: when population diversity is high, the mutation factor is decreased to accelerate convergence; when the population tends towards convergence but has not yet found a global optimum, the mutation factor is increased to escape local optima. After multiple generations of evolution, the optimal tracking angle set is obtained, which maximizes the energy harvesting efficiency of the solar panels under predicted illumination conditions.

[0067] The calculated optimal tracking angle set is input into the dual-axis drive device to perform solar panel attitude adjustment. The dual-axis drive device consists of two independent stepper motors and reducers, which control the rotation of the horizontal and pitch axes respectively. The angle difference between the current attitude and the target attitude is calculated, and then a smooth motion trajectory is generated based on the preset maximum angular velocity and acceleration to avoid impact on the mechanical structure caused by violent movements. During attitude adjustment, surface temperature distribution data is collected in real time by an array of infrared temperature sensors distributed on the surface of the solar panel to form a temperature thermal map. When an abnormal temperature distribution is detected (such as a local temperature that is significantly lower than the surrounding area), it is determined that there may be an obstruction affecting the illumination. The position, shape, and movement characteristics of the obstruction are calculated using temperature gradient field analysis, and the gradient vector field of the temperature distribution is calculated to identify gradient discontinuities as the boundaries of the obstruction. Based on the temperature distribution changes in continuous time frames, the movement speed and direction of the obstruction area are calculated. Combining geographical location, time information, and layered illumination prediction sequences, the future movement trajectory of the obstruction (such as clouds, tree shadows, etc.) is predicted to generate a dynamic obstacle avoidance scheme.

[0068] The solar panel's attitude is adjusted according to a dynamic obstacle avoidance scheme to avoid predicted shading areas and orient the solar panel towards the optimal sunlight direction. After adjustment, the output power data of the solar panel continues to be collected, and the actual energy harvesting efficiency is recorded. This data is used as input data for the next round of differential evolution calculations, forming a closed-loop feedback optimization mechanism. The above process is repeated periodically (usually at 5-15 minute intervals) to continuously optimize the solar panel's attitude, adapt to constantly changing environmental conditions, and achieve maximum energy harvesting efficiency.

[0069] For example, a solar power station located at 35 degrees north latitude and 114 degrees east longitude is equipped with the intelligent tracking and regulation system of this invention. On a typical sunny summer day, it begins operation at 6:00 AM. It obtains geographical location and current time information via a GPS module; simultaneously, it collects light intensity data through five light sensors distributed at the four corners and center of the solar panels. The current average light intensity is 320 W / m². 2The solar panel outputs 2.6kW. Empirical Mode Decomposition (EMD) was performed on the illumination data collected over the past 24 hours, extracting three main intrinsic mode functions (IMFs): the first represents rapid changes (period approximately 10-30 minutes), possibly corresponding to short-term shading from small clouds; the second represents medium-term changes (period approximately 6 hours), corresponding to diurnal variations in solar altitude angle; and the third represents long-term trends, corresponding to seasonal variations in solar intensity. Based on the characteristics of these IMFs, a hierarchical illumination prediction sequence for the next 6 hours was generated. A differential evolution algorithm was executed using historical power data and the predicted illumination sequence. The initial population contained 50 individuals, each representing a set of possible horizontal and vertical angles. After 30 generations of evolution, the current optimal angle set was calculated to be: a horizontal angle of 75 degrees (east-southeast) and a vertical angle of 42 degrees. This angle set was sent to a dual-axis drive unit, and the motor smoothly rotated the solar panel from the initial attitude (45 degrees horizontal angle, 30 degrees vertical angle) to the target attitude. During the adjustment process, the temperature sensor detected that the temperature in the upper right corner of the panel was significantly lower than other areas (temperature difference exceeding 15°C). Analysis of the temperature gradient field indicated that a cloud might be moving from west to east, partially blocking sunlight. Combining current wind speed and historical cloud movement data, it was predicted that the obstruction would move out of the solar panel area in approximately 3 minutes. Therefore, the decision was made to maintain the current optimal angle without further adjustment. Three minutes later, the obstruction moved out as predicted, and the solar panel fully received sunlight, increasing its output power to 3.2kW. This actual power data was recorded as input for the next round of differential evolution algorithm to continuously optimize the tracking strategy. Through this continuous environmental awareness and dynamic adjustment, the solar panel is kept in the optimal receiving position throughout the entire sunshine period, improving the daily energy harvesting efficiency compared to fixed-installation solar panels.

[0070] In one optional implementation, empirical mode decomposition is performed on the illumination intensity data to obtain intrinsic components, and features are extracted based on the frequency characteristics of the intrinsic components to generate a hierarchical illumination prediction sequence, including:

[0071] The light intensity data is segmented, and the maximum and minimum points of each segment are calculated. The upper and lower envelopes are generated by cubic spline interpolation.

[0072] Calculate the mean values ​​of the upper envelope and the lower envelope, and subtract the mean values ​​from the illumination intensity data to obtain the initial residual signal;

[0073] Repeatedly extract time-frequency features and calculate decomposition quality scores from the initial residual signal, and obtain new residual signals through envelope extraction and mean calculation until the intrinsic components are obtained;

[0074] The inherent components are sorted by frequency. The inherent component with the lowest frequency is used as the trend component to calculate the rate of change of the signal to obtain the change characteristics. The abrupt change point position and depth of the inherent component that is adapted to the movement frequency of the obstruction are used as the obstruction characteristics. The fluctuation period and amplitude of the remaining high-frequency inherent components are used as the disturbance characteristics.

[0075] The change prediction value is calculated based on the change characteristics, the periodic prediction value is calculated based on the occlusion characteristics, and the fluctuation prediction value is calculated based on the disturbance characteristics. The change prediction value, the periodic prediction value, and the fluctuation prediction value are superimposed to generate a layered illumination prediction sequence.

[0076] In one specific implementation, the light intensity data is segmented, with each segment consisting of 200 data points. Within each segment, maxima and minima are identified by comparing adjacent data points. For example, in the first segment [850, 920, 980, 1010, 995...], a data point is considered a maximum if its value is greater than its immediate neighbors, and a minimum if its value is less. After identifying all maxima, a cubic spline interpolation method is used to generate the upper envelope. Specifically, the coordinates of all maxima are used as control points, and a series of cubic polynomial coefficients are solved to ensure a smooth curve connecting all control points. Similarly, cubic spline interpolation is performed on all minima to generate the lower envelope.

[0077] Calculate the mean values ​​of the upper and lower envelopes at each time point. For example, at time point t = 50s, if the upper envelope value is 1200 and the lower envelope value is 800, then the mean value is 1000. Subtract this mean sequence from the original light intensity data to obtain the initial residual signal. For example, if the original data point value is 1100 and the mean value at the corresponding time point is 1000, then the residual is 100.

[0078] The initial residual signal is repeatedly subjected to time-frequency feature extraction and decomposition quality score calculation. Time-frequency feature extraction includes re-identifying the maxima and minima in the residual signal, generating new upper and lower envelopes, calculating a new mean, and subtracting this mean from the original residual signal to obtain the new residual signal. The decomposition quality score is evaluated by calculating the ratio of the energy of the new residual signal to the energy of the original signal. When the energy ratio is less than a preset threshold of 0.05, the current residual signal is determined to have become an intrinsic component. For example, if the root mean square value of the new residual signal is 25 and the root mean square value of the original signal is 500, the energy ratio is 0.0025, which meets the condition of less than 0.05. At this time, the residual signal is identified as an intrinsic component and saved. The above process is repeated until the final residual signal changes minimally or the preset maximum number of decompositions (15 times) is reached, thereby obtaining a series of intrinsic components and the final residual.

[0079] The obtained intrinsic components are sorted by frequency, and their dominant frequency is determined by calculating the zero-crossing rate (the number of times the signal changes from positive to negative or vice versa per unit time). For example, if an intrinsic component crosses zero 1000 times in 10000 seconds of data, its zero-crossing rate is 0.1 Hz. The intrinsic component with the lowest frequency (the one with the lowest zero-crossing rate) is identified as the trend component, and its average rate of change over adjacent time windows is calculated as the change characteristic. For example, if the trend component has an average value of 800 from 8:00 to 9:00 and an average value of 880 from 9:00 to 10:00, then its rate of change is 10% per hour.

[0080] Identify intrinsic components that correspond to the frequency of the occlusion's movement, typically those in the range of 0.01 Hz to 0.05 Hz. Look for abrupt changes within these components by calculating the first-order difference of the signal and setting a threshold (e.g., three times the standard deviation). When the difference exceeds the threshold, the point is marked as an abrupt change, and its location and depth are recorded. For example, an abrupt change detected at time t = 3600 s with a depth of -200 indicates a significant decrease in illumination at that moment. The location and depth of these abrupt changes constitute the occlusion characteristics.

[0081] The remaining high-frequency intrinsic components (with a zero-crossing rate greater than 0.05 Hz) represent environmental disturbances. The oscillation period (the average time interval between adjacent maxima) and amplitude (the average difference between a maximum and an adjacent minimum) of these components are calculated as disturbance characteristics. For example, a certain high-frequency component has an average oscillation period of 10 seconds and an average amplitude of 50.

[0082] A hierarchical illumination prediction sequence is generated based on the extracted three types of features. Long-term trend changes are predicted based on variation features; for example, if the current time is 10:00, the illumination intensity is 1000, and the change rate is 10% / hour, then the predicted change value at 11:00 is 1100. Periodic changes are predicted based on occlusion features; by analyzing the temporal patterns of abrupt change points in historical data, potential future occlusion events are predicted. For example, if a decrease in illumination depth of approximately -150 is detected between 14:00 and 15:00 each day, this periodic prediction value is added to the corresponding time period in the prediction sequence. Random fluctuation components are generated based on perturbation features; according to the fluctuation period and amplitude characteristics of high-frequency components in historical data, a stochastic process is used to simulate and generate fluctuation prediction values; for example, a random fluctuation of ±25 is superimposed on the base prediction value.

[0083] The variable prediction value, periodic prediction value, and fluctuation prediction value are added together according to the corresponding time points to generate the final layered illumination prediction sequence. For example, at the prediction time point t = 43200s (i.e., 12:00), if the variable prediction value is 1200, the periodic prediction value is -100 (indicating that there is shading in the prediction), and the fluctuation prediction value is +20, then the final predicted illumination intensity at that time is 1120.

[0084] In one optional implementation, the time-frequency feature extraction and decomposition quality score calculation are repeatedly performed on the initial residual signal, and a new residual signal is obtained through envelope extraction and mean calculation, until the intrinsic components are obtained, including:

[0085] The initial residual signal is segmented according to a sliding window, and the time-domain statistics and frequency-domain power spectrum within each window are extracted to construct a time-frequency feature matrix. The time-frequency feature matrix is ​​then input into a pre-trained deep neural network, which identifies nonlinear feature patterns in the initial residual signal through a multi-layer attention mechanism and outputs a feature similarity distribution map.

[0086] The stationarity index, spectral entropy, and fractal dimension of the residual signal are calculated based on the feature similarity distribution map. Fuzzy membership functions are set for the stationarity index, spectral entropy, and fractal dimension. Fuzzy rule inference is performed to obtain the current decomposition quality score.

[0087] The current decomposition quality score and time-frequency feature matrix are stored in a dynamic memory pool. A reinforcement learning method is used to establish a mapping relationship between the score and the time-frequency feature matrix. Based on the mapping relationship, the quality improvement value of the next round of decomposition is calculated, and the current termination probability value is generated.

[0088] When the current termination probability value is greater than the adaptive threshold, the decomposition process is stopped and all the obtained intrinsic components are output; when the current termination probability value is less than or equal to the adaptive threshold, the fuzzy rule weights are updated, and a new residual signal is obtained through envelope extraction and mean calculation; the adaptive threshold is updated based on all historical decomposition scores stored in the memory pool.

[0089] In one specific implementation, the initial residual signal is segmented according to a sliding window, and time-domain statistics and frequency-domain power spectrum are extracted within each window to construct a time-frequency feature matrix. The size of the sliding window is set to 5% of the original signal length, and the window overlap rate is 50%. For each window, time-domain statistics are calculated, including mean, standard deviation, kurtosis, skewness, maximum value, minimum value and their occurrence positions, and zero-crossing rate. Frequency-domain features are obtained by performing a Fast Fourier Transform on the signal within the window, extracting the amplitude and corresponding frequency of the main frequency components, as well as the frequency band energy distribution characteristics. Assuming the original signal length is 10,000 sampling points, the sliding window size is 500 points, with adjacent windows overlapping by 250 points, resulting in a total of 39 windows. Eight time-domain features and ten frequency-domain features are extracted from each window, ultimately constructing a 39×18 time-frequency feature matrix.

[0090] The time-frequency feature matrix is ​​input into a pre-trained deep neural network. This deep neural network identifies nonlinear feature patterns in the initial residual signal through a multi-layer attention mechanism and outputs a feature similarity distribution map. The deep neural network employs an encoder-decoder structure. The encoder consists of three convolutional layers, each followed by batch normalization and a ReLU activation function. The kernel sizes are 3×3, 3×3, and 5×5, with a stride of 1 for each layer, and output channels of 16, 32, and 64, respectively. The decoder also contains three deconvolutional layers, with a structure symmetrical to the encoder. A multi-head self-attention layer with four heads and a hidden layer dimension of 256 is placed between the encoder and decoder. The attention mechanism calculates the correlation between elements in the feature matrix, generating an attention weight matrix to highlight key feature patterns in the signal. The trained network processes the input time-frequency feature matrix and outputs a feature similarity distribution map, represented as a 39×39 matrix. Each element represents the similarity between features within a corresponding window, with values ​​ranging from 0 to 1.

[0091] The stationarity index, spectral entropy, and fractal dimension of the residual signal are calculated based on the feature similarity distribution map. Fuzzy membership functions are then set for the stationarity index, spectral entropy, and fractal dimension, and fuzzy rule inference is performed to obtain the current decomposition quality score. The stationarity index is obtained by calculating the average of the diagonal elements of the feature similarity distribution map. This index reflects the self-similarity of different parts of the signal; a value closer to 1 indicates a more stationary signal. The spectral entropy is obtained by calculating the probability distribution entropy of the off-diagonal elements of the feature similarity distribution map, reflecting the complexity of the signal's spectral distribution. A higher value indicates a more uniform spectral distribution. The fractal dimension is extracted from the feature similarity distribution map using box counting, reflecting the complexity of the signal structure. For these three indicators, a triangular membership function is designed to divide the stationarity index into three fuzzy sets: "low," "medium," and "high," with thresholds of 0.3 and 0.7, respectively; the spectral entropy is divided into three fuzzy sets: "narrow," "medium," and "wide," with thresholds of 0.4 and 0.8, respectively; and the fractal dimension is divided into three fuzzy sets: "simple," "medium," and "complex," with thresholds of 1.2 and 1.8, respectively. Based on these fuzzy sets, 15 fuzzy rules are designed, such as "if the stationarity index is high, the spectral entropy is narrow, and the fractal dimension is simple, then the decomposition quality is extremely high." Through fuzzy rule reasoning, the current decomposition quality score is finally obtained, with a score range from 0 to 100.

[0092] The current decomposition quality score and time-frequency feature matrix are stored in a dynamic memory pool. A reinforcement learning method is used to establish a mapping relationship between the score and the time-frequency feature matrix. Based on this mapping relationship, the quality improvement value for the next round of decomposition is calculated, generating the current termination probability value. The dynamic memory pool adopts a priority queue structure, with a capacity of the results of the most recent 20 decompositions. Each record contains the time-frequency feature matrix, the decomposition quality score, and its corresponding processing parameters. The reinforcement learning model uses a deep Q-network, with the state space being the current time-frequency feature matrix, the action space being the decomposition parameter adjustment strategy, and the reward function being the difference in quality scores between two adjacent decompositions. Model training uses an experience replay mechanism, with a batch size of 32, a learning rate of 0.001, and a discount factor of 0.95. Based on the trained model, the quality improvement value for the next round of decomposition is predicted, i.e., the expected increment in quality score after executing the optimal action. The current termination probability value is calculated as: 1 minus the quality improvement value divided by the current quality score, multiplied by the adjustment factor 0.8. For example, if the current decomposition quality score is 85 and the predicted quality improvement value for the next round of decomposition is 5, then the current termination probability value is 1 minus 5 divided by 85, and then multiplied by 0.8, resulting in approximately 0.75.

[0093] When the current termination probability value is greater than the adaptive threshold, the decomposition process stops and all obtained intrinsic components are output. When the current termination probability value is less than or equal to the adaptive threshold, the fuzzy rule weights are updated, and a new residual signal is obtained through envelope extraction and mean calculation. The adaptive threshold is updated based on all historical decomposition scores stored in the memory pool. The initial value of the adaptive threshold is set to 0.7, and it is adjusted according to the historical decomposition scores stored in the dynamic memory pool as the decomposition process progresses. Specifically, after each round of decomposition, the average growth rate of the historical scores is calculated. If the average growth rate is less than 1%, the adaptive threshold is reduced by 0.05; if the average growth rate is greater than 5%, the adaptive threshold is increased by 0.05. The threshold value is limited to the range of 0.5 to 0.9. When the termination probability value is greater than the adaptive threshold, it is determined that the benefit of continuing the decomposition is not significant, the decomposition process stops, and all obtained intrinsic components are output. If the termination probability value is less than or equal to the adaptive threshold, the decomposition continues. First, based on the quality score of this round of decomposition, the fuzzy rule weights are updated, strengthening the rule weights corresponding to high-quality decomposition and weakening the rule weights corresponding to low-quality decomposition. Then, a new residual signal is obtained through envelope extraction and mean calculation. Envelope extraction employs cubic spline interpolation. Upper and lower envelopes are constructed at extreme points, their mean is calculated, and this mean is subtracted from the original signal to obtain a new residual signal, which serves as the input for the next round of decomposition. For example, for a signal with 10,000 points, approximately 150 maxima and minima are first detected, upper and lower envelopes are constructed for each, their mean is calculated, and then subtracted from the original signal to obtain a new residual signal.

[0094] Traditional empirical mode decomposition (EMD) methods primarily extract intrinsic mode functions (IMFs) of signals through iterative sieving processes. However, they lack effective decomposition quality evaluation mechanisms and often rely on fixed termination conditions, such as the number of iterations or error thresholds, making it difficult to adapt to the characteristics of different signal types. Existing technologies typically use simple indicators such as energy ratio and correlation coefficient to evaluate decomposition quality, lacking in-depth analysis of the complex nonlinear characteristics of signals. This embodiment proposes an adaptive signal decomposition method based on deep neural networks and fuzzy reinforcement learning to construct an intelligent evaluation and decision-making mechanism for signal decomposition. It extracts nonlinear feature patterns from the residual signal through deep neural networks, evaluates decomposition quality using a fuzzy logic system, and optimizes the decomposition strategy and termination decision using reinforcement learning. The resulting improvements significantly enhance the adaptability and effectiveness of signal decomposition, especially for signals with complex modulation and non-stationary characteristics, improving decomposition accuracy and computational efficiency while reducing the risks of over-decomposition and under-decomposition, providing a more reliable foundation for signal feature extraction and pattern recognition.

[0095] like Figure 2 The diagram shown illustrates the residual signal decomposition process.

[0096] In one optional implementation, differential evolution is performed using historical output power data and hierarchical illumination prediction sequences, combined with adaptive adjustment of the mutation factor, to calculate the optimal tracking angle set, including:

[0097] Historical output power data and hierarchical illumination prediction sequences are used as input data for differential evolution. Each tracking angle group is set to include azimuth and tilt angles. Initial population individuals are generated by random sampling within a preset range. Each initial population individual includes an initial azimuth and an initial tilt angle.

[0098] The Euclidean distance between individuals in the initial population is calculated to obtain the inter-individual distance matrix. The mean and variance of the inter-individual distance matrix are statistically analyzed to determine the population clustering degree. The adjustment step size of the variation factor is calculated based on the population clustering degree. The adjustment step size is superimposed with the current variation factor to form an adaptive adjustment term for the variation factor.

[0099] The global optimal individual is selected from the initial population as the basis vector. Two different individuals are selected to calculate the difference vector. The mutation factor adaptive adjustment term is multiplied by the difference vector and superimposed on the basis vector to generate the mutated individual.

[0100] Calculate the predicted power values ​​of the mutated individuals and the initial population individuals respectively, select the individual corresponding to the maximum predicted power value to enter the next generation population, and repeat the process until the optimal tracking angle group is obtained.

[0101] In one specific implementation, historical output power data of the solar panel is acquired, including the actual output power values ​​of the solar panel at different time points and under different angle combinations. Simultaneously, real-time illumination data collected by a light intensity sensor is combined to construct a hierarchical illumination prediction sequence based on a deep learning model. The hierarchical illumination prediction sequence refers to a prediction dataset formed by combining short-term, medium-term, and long-term illumination intensity change trends through a multi-level prediction model. The historical output power data and the hierarchical illumination prediction sequence serve as input data for a differential evolution algorithm, used in the subsequent process of solving for the optimal angle combination.

[0102] Each tracking angle group is configured to include two parameters: azimuth and tilt. The azimuth represents the angle of horizontal rotation of the solar panel, ranging from 0° to 359°; the tilt represents the degree of tilt of the solar panel relative to the horizontal plane, ranging from 0° to 90°. An initial population of individuals is generated using random sampling within a preset range, with a population size of 50 individuals. Each initial individual contains two parameter values: initial azimuth and initial tilt. For example, possible parameter combinations for individual 1 are (azimuth = 120°, tilt = 35°), and for individual 2 are (azimuth = 175°, tilt = 42°), and so on, generating 50 different angle combinations as the initial population.

[0103] Calculate the Euclidean distances between individuals in the initial population to construct an inter-individual distance matrix. For any two individuals i and j, their Euclidean distance is calculated as the square root of the sum of the squares of their azimuth and tilt angle differences. For example, the Euclidean distance between individual 1 (120°, 35°) and individual 2 (175°, 42°) is 59.21. A 50×50 distance matrix is ​​formed by calculating the pairwise distances between all individuals. The mean and variance of this matrix are then used as indicators of population aggregation. A smaller mean and variance indicate higher population aggregation; a larger mean and variance indicate higher population dispersion.

[0104] The adjustment step size of the variation factor is calculated based on the population aggregation degree. When the population aggregation degree is high, the adjustment step size is increased to promote the population to explore a wider solution space; when the population dispersion degree is high, the adjustment step size is decreased to enhance local fine-grained search. In specific implementation, if the mean of the distance matrix is ​​less than 30 and the variance is less than 100, the adjustment step size is set to 0.1; if the mean is between 30 and 60 and the variance is between 100 and 300, the adjustment step size is set to 0.05; if the mean is greater than 60 or the variance is greater than 300, the adjustment step size is set to 0.02. The adjustment step size is superimposed with the current variation factor to form an adaptive adjustment term for the variation factor. The initial variation factor is set to 0.5, and the variation factor after adaptive adjustment is controlled between 0.3 and 0.9.

[0105] The globally optimal individual is selected from the initial population as the basis vector. The globally optimal individual is determined by calculating the predicted power value of each individual under the current illumination conditions, and the individual with the highest predicted power value is selected. For example, if the predicted power value of an individual (azimuth = 150°, tilt = 40°) is 850 watts, which is the highest value in the current population, then this individual is selected as the basis vector. Subsequently, two different individuals are randomly selected, and the difference vector between these two individuals is calculated. For example, if individuals (azimuth = 130°, tilt = 35°) and individuals (azimuth = 170°, tilt = 45°) are selected, their difference vector is (40°, 10°).

[0106] The adaptive adjustment term of the mutation factor is multiplied by the difference vector and then superimposed onto the basis vectors to generate the mutated individual. Assuming the current adaptive adjustment term of the mutation factor is 0.6, the difference vector multiplied by the mutation factor is (24°, 6°), which, when superimposed onto the basis vector (150°, 40°), yields the mutated individual (174°, 46°). Boundary checks are performed on the mutated individuals to ensure that the azimuth angle is within the range of 0° to 359° and the tilt angle is within the range of 0° to 90°; if these ranges are exceeded, corrections are made.

[0107] The predicted power values ​​of the mutant individuals and the initial population individuals under the current illumination conditions are calculated using an illumination model. The model considers factors such as solar altitude angle, azimuth angle, cloud cover, and atmospheric transparency to accurately simulate the solar panel's light reception under different angle combinations. For example, the predicted power value of the mutant individual (174°, 46°) is 872 watts, while the predicted power value of the optimal individual in the initial population (150°, 40°) is 850 watts. In this case, the mutant individual is selected to enter the next generation population.

[0108] Repeat the above steps, including calculating population aggregation, adaptively adjusting the mutation factor, generating mutated individuals, and selecting better individuals, until the termination condition is met. The termination condition can be set to reach the maximum number of iterations (e.g., 100 times) or the change in the optimal solution over 10 consecutive generations being less than a preset threshold (e.g., 1%). The final optimal tracking angle set is the combination of azimuth and tilt angles that maximizes the output power of the solar panel under the current illumination conditions.

[0109] In one alternative implementation, it further includes:

[0110] Based on the initial population individuals, the initial value and mapping equation for iterative calculation are determined. The initial value is substituted into the mapping equation to obtain the first iteration result. The first iteration result is used as input and substituted into the mapping equation again to obtain the second iteration result. The iteration is repeated until the number of iterations reaches the size of the initial population. Each iteration result is a chaotic value between zero and one.

[0111] The adaptive coefficient is determined based on the number of generations of the current population. The adaptive coefficient decreases as the number of generations increases. The adaptive coefficient is multiplied by the chaotic value to determine the chaotic perturbation term.

[0112] The chaotic perturbation term is allocated to the azimuth angle and tilt angle according to a preset ratio to perturb the initial population individuals, thereby obtaining the perturbed population individuals.

[0113] When calculating the new generation population, based on the preset number of iterations, the relationship between the increase in the optimal predicted output power value of multiple consecutive generations of population and the preset magnitude threshold is determined, and the adaptive coefficient is adjusted according to the determination result.

[0114] In one specific implementation, during the solar panel illumination tracking optimization process, the initial values ​​and mapping equation required for chaotic mapping are determined based on the generated initial population. In this embodiment, the Logistic mapping is used as the mathematical model for the chaotic system. The initial value is selected as a random number between 0.2 and 0.8, avoiding the special points of 0, 0.25, 0.5, 0.75, and 1 in the Logistic mapping to ensure chaos. Taking an initial population size of 50 as an example, an initial value of 0.38 is randomly selected. The mapping equation is four times the current value multiplied by one minus the current value. This equation exhibits typical chaotic characteristics, capable of generating seemingly random but deterministic sequences between zero and one.

[0115] Substituting the initial value of 0.38 into the mapping equation yields the first iteration result of 0.9424. Using this result as input, substituting it into the mapping equation again yields the second iteration result of 0.2174. Repeating this process, subsequent iteration results are obtained: 0.6807, 0.8697, 0.4539, 0.9911, 0.0352, etc. This iteration continues until 50 chaotic values ​​are obtained, matching the initial population size. These chaotic values, all between zero and one, are unpredictable and ergodic, introducing diverse perturbations into the population.

[0116] The adaptive coefficient is determined based on the current generation number of the population to dynamically adjust the intensity of chaotic perturbation. In the early stages of optimization, a larger perturbation is beneficial for global search; as optimization progresses, gradually reducing the perturbation intensity helps the algorithm converge to an exact solution. The adaptive coefficient is calculated using an exponential decay method, with an initial value of 0.5. For each additional generation, the adaptive coefficient is multiplied by 0.98. For example, the adaptive coefficient is 0.5 for generation 1, 0.4057 for generation 10, 0.2031 for generation 50, and 0.0913 for generation 100. The chaotic perturbation term is obtained by multiplying the adaptive coefficient by the chaotic value. Taking generation 10 as an example, if the corresponding chaotic value is 0.7586, then the chaotic perturbation term is 0.4057 multiplied by 0.7586, which equals 0.3078.

[0117] The chaotic perturbation term is allocated to the azimuth and tilt angles according to a preset ratio, perturbing the initial population individuals to obtain a perturbed population. Considering the different ranges of azimuth and tilt angles (azimuth from 0° to 359°, tilt angle from 0° to 90°), the perturbation allocation ratio is set to 4:1. Taking a chaotic perturbation term of 0.3078 as an example, the perturbation allocated to the azimuth angle is 0.3078 multiplied by 0.8, which equals 0.2462, and the perturbation allocated to the tilt angle is 0.3078 multiplied by 0.2, which equals 0.0616. Converting the perturbation values ​​to angular changes, the azimuth perturbation is 0.2462 multiplied by 360°, which equals 88.63°, and the tilt angle perturbation is 0.0616 multiplied by 90°, which equals 5.54°.

[0118] An initial population individual (azimuth = 150°, tilt = 40°) is perturbed to obtain a perturbed individual (azimuth = 238.63°, tilt = 45.54°). Boundary checks are performed on the perturbed angles to ensure the azimuth is within the range of 0° to 359° and the tilt is within the range of 0° to 90°. If these ranges are exceeded, a reflection strategy is used. For example, if the perturbed tilt becomes 95°, exceeding the 90° upper limit, it is adjusted to 85° (i.e., 90° - (95° - 90°)). The perturbed population participates in subsequent selection operations together with the original population, increasing population diversity and avoiding local optima.

[0119] When calculating the next generation of the population, based on a preset number of iterations, the relationship between the increase in the optimal predicted output power value of consecutive generations and a preset amplitude threshold is determined, and the adaptive coefficient is dynamically adjusted according to the determination result. The preset number of iterations is 10 generations, and the preset amplitude threshold is set to 1%. If the increase in the optimal predicted output power value of consecutive 10 generations is less than 1%, it indicates that the algorithm may be trapped in a local optimum or close to convergence. In this case, the adaptive coefficient is increased by 50%, but not exceeding the initial value of 0.5, to enhance the perturbation strength and help the algorithm escape local optima. If the increase in the optimal predicted output power value of consecutive 10 generations is greater than 5%, it indicates that the algorithm is in a fast convergence phase. In this case, the adaptive coefficient is decreased by 20% to accelerate the convergence speed.

[0120] For example, assuming the optimal predicted output power values ​​for generations 40 to 49 are 852 watts, 854 watts, 855 watts, 856 watts, 856 watts, 857 watts, 857 watts, 858 watts, 858 watts, and 859 watts respectively, the calculated improvement is (859 - 852) / 852 = 0.82%, which is less than the preset threshold of 1%. Therefore, it is determined that the algorithm may be trapped in a local optimum. The adaptive coefficient for generation 50 is increased by 50% from the original 0.2031 to 0.3047 to strengthen the perturbation. After the perturbation, the optimal predicted output power value for generation 50 increases to 865 watts, successfully escaping the local optimum.

[0121] By iteratively executing the above steps, including generating a perturbation sequence through chaotic mapping, adaptively adjusting the perturbation intensity, perturbing individuals in the population, and dynamically adjusting the adaptive coefficients, the process continues until a termination condition is met. The termination condition can be set to reach a maximum of 200 iterations, or the optimal solution variation must be less than 0.5% for 30 consecutive iterations. The final optimal tracking angle set is used as the actual control parameters for the solar panel, achieving precise optimization of light tracking.

[0122] Traditional solar panel tracking optimization methods mainly rely on deterministic algorithms or simple random searches, which are prone to getting trapped in local optima and have poor adaptability to different lighting conditions. The method in this embodiment combines chaos theory with differential evolution algorithms, using the unpredictability and ergodicity of chaotic systems to introduce nonlinear perturbations into the evolutionary process, and adopts an adaptive adjustment strategy to dynamically control the perturbation intensity, significantly enhancing the algorithm's global search capability and local refinement capability.

[0123] like Figure 3 As shown, a flowchart illustrating the solar tracking angle optimization process using differential evolution and chaotic perturbations is presented.

[0124] In one optional implementation, when a temperature anomaly is detected, the movement characteristics of the obstruction are calculated using temperature gradient field analysis, and a dynamic obstacle avoidance scheme is generated by combining this with an illumination prediction sequence, including:

[0125] Collect temperature distribution data on the surface of the solar panel, calculate the mean and standard deviation of the temperature distribution data at different time points, and obtain the temperature distribution baseline value under normal operating conditions;

[0126] The temperature distribution data is subtracted from the temperature distribution benchmark value to obtain the temperature difference data, and the temperature anomaly area is determined based on the time change of the temperature difference data.

[0127] Boundary extraction is performed on the temperature anomaly region, the temperature gradient value of the temperature anomaly region is calculated, adjacent temperature gradient values ​​are clustered, and the centroid coordinates and area of ​​the clustered region are calculated to obtain the initial spatial position of the occlusion.

[0128] Calculate the temperature gradient values ​​in the horizontal and vertical directions of the temperature anomaly region to obtain the temperature gradient field, and record the gradient change values ​​of the temperature gradient field over a continuous time period.

[0129] The direction of the largest temperature change is extracted from the temperature gradient field as the main direction. Feature points are selected along the main direction, and the position changes of the feature points at different times are tracked to determine the contour change of the occluder. The moving direction and moving speed of the occluder are calculated based on the contour change.

[0130] The location coordinates of the obstruction in the future time period are calculated based on the direction and speed of movement, and the location coordinates are combined with the layered illumination prediction sequence to calculate the obstruction area.

[0131] Within a preset time window, the obstacle avoidance adjustment angle is calculated based on the occlusion area. The obstacle avoidance adjustment angle is then executed in segments. After each segment is executed, the obstacle avoidance adjustment angle for the next segment is calculated based on the new occlusion area, until the entire obstacle avoidance process is completed.

[0132] In one specific implementation, temperature distribution data on the surface of the solar panel is collected at a frequency of once per minute. Each collected temperature data point forms a two-dimensional matrix, the size of which corresponds to the size of the solar panel. For example, for a standard 1.6m × 1.0m solar panel, it can be divided into a 160 × 100 array of temperature collection points, with each point recording a temperature value. Time-series statistics are performed on the collected temperature distribution data to calculate the mean and standard deviation of the temperature at the same time point over seven consecutive days under normal sunlight conditions for each collection point. For example, for the collection point at coordinates (50, 30), the normal mean temperature at 10:00 AM is 42.5℃, with a standard deviation of 1.2℃. These statistical data constitute the baseline values ​​for the temperature distribution of the solar panel at different time points, serving as a reference standard for subsequent anomaly detection.

[0133] The temperature difference data is obtained by subtracting the real-time temperature distribution data from the baseline temperature distribution value at the corresponding time point. Under normal circumstances, the temperature difference should be close to zero or fluctuate within the standard deviation range. When the solar panel is shaded, the temperature in the shaded area will be significantly lower than the baseline value, forming a negative temperature difference. For example, the measured temperature at a certain time coordinate (80, 45) is 35.2℃, while the baseline temperature at that time is 46.8℃, resulting in a temperature difference of -11.6℃, which far exceeds the standard deviation range and is identified as a temperature anomaly. By comparing the changes in temperature difference values ​​at multiple consecutive time points, areas with persistent anomalies are identified as temperature anomaly areas. The temperature change threshold is set at 3℃; when the temperature difference value in a certain area changes by more than 3℃ within 3 minutes, it is identified as an anomaly area.

[0134] Boundary extraction was performed on temperature anomaly regions. The Canny edge detection algorithm was used to process the temperature difference matrix, extracting points of sudden temperature changes to form boundary contours. The temperature difference between each point within the temperature anomaly region and its neighbors was calculated to obtain the temperature gradient value. The temperature gradient value represents the rate of temperature change per unit distance, expressed in °C / cm. Adjacent points with similar temperature gradient values ​​were clustered using the density-based DBSCAN algorithm, with a similarity threshold of 2 °C / cm and a minimum cluster size of 5 points. The centroid coordinates and area of ​​each cluster were calculated. The centroid coordinates were calculated by weighting the coordinates of all points within the region, with the weights being the temperature gradient values ​​of each point. The area of ​​the region was obtained by multiplying the number of cluster points by the area per unit sampling point. For example, a cluster region might contain 85 sampling points, each representing 1 cm. 2 If the area is 85cm², then the area of ​​this region is 85cm². 2 The centroid coordinates are (92.3, 57.6), and these parameters constitute the initial spatial position information of the occluder.

[0135] The temperature gradient values ​​in the horizontal and vertical directions of the temperature anomaly region are calculated to construct a two-dimensional temperature gradient field. The horizontal gradient is obtained by calculating the temperature difference between adjacent points in the same row, and the vertical gradient is obtained by calculating the temperature difference between adjacent points in the same column. The gradient values ​​in the two directions are combined to form a gradient vector, where the magnitude of the vector represents the rate of temperature change and the direction represents the steepest temperature change. The changes in the temperature gradient field are recorded over a continuous time period, with gradient values ​​recorded at 5-second intervals for 12 consecutive time points, forming a gradient change sequence.

[0136] In the temperature gradient field, the direction with the largest temperature change is extracted as the principal direction. Specifically, this is done by calculating the direction histogram of all gradient vectors in the gradient field and selecting the direction interval with the highest frequency as the principal direction. Feature points are selected along the principal direction, with the selection criterion being the point with the largest local gradient value. One feature point is selected every 5 cm, for a total of 10-15 feature points. An optical flow algorithm is used to track the position changes of these feature points at different times, recording the position coordinates at five consecutive time points and calculating the displacement vector. Based on the displacement information of the feature points, the least squares method is used to fit the overall motion pattern of the occluder, determining the change law of the occluder's contour. The moving direction and speed of the occluder are calculated based on the contour changes. The moving direction is expressed as an angle value, ranging from 0° to 359°, and the moving speed is in cm / min. For example, the moving direction of an occluder is 225° (southwest), and the moving speed is 18 cm / min.

[0137] Based on the direction and speed of the obstruction's movement, a linear prediction model is used to calculate the obstruction's position coordinates in future time periods. The prediction duration is set to 10 minutes, with position sequences generated at 30-second intervals. For example, if the centroid of the obstruction is currently at time t0 (92.3, 57.6), its predicted position at time t0+5 minutes is (47.3, 12.6). The predicted position coordinates are combined with a layered illumination prediction sequence to calculate the obstruction area and its impact on the solar panel's output power. The layered illumination prediction sequence contains illumination intensity predictions at different time granularities. Combining the obstruction's position and shape, the effective obstruction area is calculated using a Monte Carlo integration method, expressed in square centimeters.

[0138] Within a preset 5-minute time window, the obstacle avoidance adjustment angle is calculated based on the obstruction area. The calculation considers the current angle of the solar panel, the movement trajectory of the obstruction, and the direction of sunlight, aiming to minimize the obstruction area. The adjustment angle includes two parameters: azimuth and tilt angle. The azimuth adjustment range is 0° to 359°, and the tilt angle range is 0° to 90°. The obstacle avoidance adjustment employs a segmented execution strategy, dividing the 5-minute time window into five 1-minute segments. Each segment adjusts an angle no greater than 15° to ensure stable motor load. After each segment is completed, temperature distribution data is re-collected, the obstruction's position and movement parameters are updated, and the obstacle avoidance adjustment angle for the next segment is calculated. For example, the first segment adjusts the azimuth angle by +12° and the tilt angle by -5°; after completion, the obstruction's movement speed changes to 15 cm / min, and its direction changes to 240°. Based on this, the second segment is calculated with an azimuth angle of +8° and a tilt angle of -3°. This dynamic adjustment method enables real-time tracking and obstacle avoidance of moving obstructions until the entire obstacle avoidance process is completed, allowing the solar panel to return to its optimal light-receiving state.

[0139] Traditional solar panel tracking systems primarily rely on geometric calculations of the sun's position, lacking the ability to perceive and process shading. They fail to respond effectively to dynamic obstructions such as tree branches, floating objects, or temporary structures, leading to a significant decrease in power generation efficiency. Existing shading detection methods mostly employ photoelectric sensor arrays or camera-based visual recognition; the former suffers from low resolution and high cost, while the latter exhibits poor accuracy and high computational demands under strong sunlight. This embodiment utilizes the surface temperature distribution characteristics of the solar panel for shading detection, combining temperature gradient field analysis and feature point tracking technology to predict the trajectory of shading objects, achieving accurate identification and prediction of dynamic shading. Through a segmented obstacle avoidance strategy, it not only effectively avoids the impact of shading but also maintains stable system operation and reduces mechanical wear.

[0140] The solar panel illumination tracking optimization system based on artificial intelligence, as described in this embodiment of the invention, includes:

[0141] The first unit is used to collect geographical location information and time information of the solar panel installation location, and to obtain light intensity data and current output power data of the solar panel through multiple light sensors;

[0142] The second unit is used to perform empirical mode decomposition on the illumination intensity data to obtain intrinsic components, extract features based on the frequency characteristics of the intrinsic components, and generate a hierarchical illumination prediction sequence.

[0143] The third unit is used to perform differential evolution operations using historical output power data and hierarchical illumination prediction sequences, combined with adaptive adjustment of mutation factors, to calculate the optimal tracking angle group;

[0144] The fourth unit is used to input the optimal tracking angle set into the dual-axis drive device to perform solar panel attitude adjustment, collect surface temperature distribution in real time, and when an abnormal temperature is detected, use temperature gradient field analysis to calculate the movement characteristics of the obstruction, and combine the layered illumination prediction sequence to generate a dynamic obstacle avoidance scheme.

[0145] The fifth unit is used to adjust the solar panel attitude according to the dynamic obstacle avoidance scheme and collect the adjusted output power data as input data for the next round of differential evolution calculation.

[0146] A third aspect of the present invention provides an electronic device, comprising:

[0147] processor;

[0148] Memory used to store processor-executable instructions;

[0149] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0150] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0151] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based method for optimizing solar panel illumination tracking, characterized in that, include: Collect geographical location and time information of the solar panel installation location, and obtain light intensity data and current output power data of the solar panel through multiple light sensors; Empirical mode decomposition (EMD) is performed on the illumination intensity data to obtain intrinsic components. Features are extracted based on the frequency characteristics of the intrinsic components to generate a hierarchical illumination prediction sequence, including: The light intensity data is segmented, and the maximum and minimum points of each segment are calculated. The upper and lower envelopes are generated by cubic spline interpolation. Calculate the mean values ​​of the upper envelope and the lower envelope, and subtract the mean values ​​from the illumination intensity data to obtain the initial residual signal; Repeatedly extract time-frequency features and calculate decomposition quality scores from the initial residual signal, and obtain new residual signals through envelope extraction and mean calculation, until the intrinsic components are obtained, including: The initial residual signal is segmented according to a sliding window, and the time-domain statistics and frequency-domain power spectrum of each window are extracted to construct a time-frequency feature matrix. The time-frequency feature matrix is ​​input into a pre-trained deep neural network, which identifies nonlinear feature patterns in the initial residual signal through a multi-layer attention mechanism and outputs a feature similarity distribution map. The stationarity index, spectral entropy, and fractal dimension of the residual signal are calculated based on the feature similarity distribution map. Fuzzy membership functions are set for the stationarity index, spectral entropy, and fractal dimension. Fuzzy rule inference is performed to obtain the current decomposition quality score. The current decomposition quality score and time-frequency feature matrix are stored in a dynamic memory pool. A reinforcement learning method is used to establish a mapping relationship between the score and the time-frequency feature matrix. Based on the mapping relationship, the quality improvement value of the next round of decomposition is calculated, and the current termination probability value is generated. When the current termination probability value is greater than the adaptive threshold, the decomposition process is stopped and all the obtained intrinsic components are output. When the current termination probability value is less than or equal to the adaptive threshold, the fuzzy rule weights are updated, and a new residual signal is obtained by envelope extraction and mean calculation. The adaptive threshold is updated based on all historical decomposition scores stored in the memory pool; The inherent components are sorted by frequency. The inherent component with the lowest frequency is used as the trend component to calculate the rate of change of the signal to obtain the change characteristics. The abrupt change point position and depth of the inherent component that is adapted to the movement frequency of the obstruction are used as the obstruction characteristics. The fluctuation period and amplitude of the remaining high-frequency inherent components are used as the disturbance characteristics. The change prediction value is calculated based on the change characteristics, the periodic prediction value is calculated based on the occlusion characteristics, and the fluctuation prediction value is calculated based on the disturbance characteristics. The change prediction value, the periodic prediction value, and the fluctuation prediction value are superimposed to generate a layered illumination prediction sequence. Differential evolution operation is performed using historical output power data and layered illumination prediction sequence, combined with adaptive adjustment of mutation factor, to calculate the optimal tracking angle group; The optimal tracking angle group is input into the dual-axis drive device to perform solar panel attitude adjustment and collect surface temperature distribution in real time. When an abnormal temperature is detected, the temperature gradient field analysis is used to calculate the movement characteristics of the obstruction and generate a dynamic obstacle avoidance scheme by combining the layered illumination prediction sequence. The solar panel attitude is adjusted according to the dynamic obstacle avoidance scheme, and the adjusted output power data is collected as input data for the next round of differential evolution calculation.

2. The method according to claim 1, characterized in that, Differential evolution operations were performed using historical output power data and layered illumination prediction sequences, combined with adaptive adjustment of mutation factors, to calculate the optimal tracking angle set, which includes: Historical output power data and hierarchical illumination prediction sequences are used as input data for differential evolution. Each tracking angle group is set to include azimuth and tilt angles. Initial population individuals are generated by random sampling within a preset range. Each initial population individual includes an initial azimuth and an initial tilt angle. The Euclidean distance between individuals in the initial population is calculated to obtain the inter-individual distance matrix. The mean and variance of the inter-individual distance matrix are statistically analyzed to determine the population clustering degree. The adjustment step size of the variation factor is calculated based on the population clustering degree. The adjustment step size is superimposed with the current variation factor to form an adaptive adjustment term for the variation factor. The global optimal individual is selected from the initial population as the basis vector. Two different individuals are selected to calculate the difference vector. The mutation factor adaptive adjustment term is multiplied by the difference vector and superimposed on the basis vector to generate the mutated individual. Calculate the predicted power values ​​of the mutated individuals and the initial population individuals respectively, select the individual corresponding to the maximum predicted power value to enter the next generation population, and repeat the process until the optimal tracking angle group is obtained.

3. The method according to claim 2, characterized in that, Also includes: Based on the initial population individuals, the initial value and mapping equation for iterative calculation are determined. The initial value is substituted into the mapping equation to obtain the first iteration result. The first iteration result is used as input and substituted into the mapping equation again to obtain the second iteration result. The iteration is repeated until the number of iterations reaches the size of the initial population. Each iteration result is a chaotic value between zero and one. The adaptive coefficient is determined based on the number of generations of the current population. The adaptive coefficient decreases as the number of generations increases. The adaptive coefficient is multiplied by the chaotic value to determine the chaotic perturbation term. The chaotic perturbation term is allocated to the azimuth angle and tilt angle according to a preset ratio to perturb the initial population individuals, thereby obtaining the perturbed population individuals. When calculating the new generation population, based on the preset number of iterations, the relationship between the increase in the optimal predicted output power value of the population over multiple generations and the preset magnitude threshold is determined, and the adaptive coefficient is adjusted according to the determination result.

4. The method according to claim 1, characterized in that, When an abnormal temperature is detected, the movement characteristics of obstructions are calculated using temperature gradient field analysis, and a dynamic obstacle avoidance scheme is generated by combining it with the illumination prediction sequence, including: Collect temperature distribution data on the surface of the solar panel, calculate the mean and standard deviation of the temperature distribution data at different time points, and obtain the temperature distribution baseline value under normal operating conditions; The temperature distribution data is subtracted from the temperature distribution benchmark value to obtain the temperature difference data, and the temperature anomaly area is determined based on the time change of the temperature difference data. Boundary extraction is performed on the temperature anomaly region, the temperature gradient value of the temperature anomaly region is calculated, adjacent temperature gradient values ​​are clustered, and the centroid coordinates and area of ​​the clustered region are calculated to obtain the initial spatial position of the occlusion. Calculate the temperature gradient values ​​in the horizontal and vertical directions of the temperature anomaly region to obtain the temperature gradient field, and record the gradient change values ​​of the temperature gradient field over a continuous time period. The direction of the largest temperature change is extracted from the temperature gradient field as the main direction. Feature points are selected along the main direction, and the position changes of the feature points at different times are tracked to determine the contour change of the occluder. The moving direction and moving speed of the occluder are calculated based on the contour change. The location coordinates of the obstruction in the future time period are calculated based on the direction and speed of movement, and the location coordinates are combined with the layered illumination prediction sequence to calculate the obstruction area. Within a preset time window, the obstacle avoidance adjustment angle is calculated based on the occlusion area. The obstacle avoidance adjustment angle is then executed in segments. After each segment is executed, the obstacle avoidance adjustment angle for the next segment is calculated based on the new occlusion area, until the entire obstacle avoidance process is completed.

5. An artificial intelligence-based solar panel illumination tracking optimization system, used to implement the method of any one of claims 1-4, characterized in that, include: The first unit is used to collect geographical location information and time information of the solar panel installation location, and to obtain light intensity data and current output power data of the solar panel through multiple light sensors; The second unit is used to perform empirical mode decomposition on the illumination intensity data to obtain intrinsic components, extract features based on the frequency characteristics of the intrinsic components, and generate a hierarchical illumination prediction sequence. The third unit is used to perform differential evolution operations using historical output power data and hierarchical illumination prediction sequences, combined with adaptive adjustment of mutation factors, to calculate the optimal tracking angle group; The fourth unit is used to input the optimal tracking angle set into the dual-axis drive device to perform solar panel attitude adjustment, collect surface temperature distribution in real time, and when an abnormal temperature is detected, use temperature gradient field analysis to calculate the movement characteristics of the obstruction, and combine the layered illumination prediction sequence to generate a dynamic obstacle avoidance scheme. The fifth unit is used to adjust the solar panel attitude according to the dynamic obstacle avoidance scheme and collect the adjusted output power data as input data for the next round of differential evolution calculation.

6. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 4.

Images