Mountain photovoltaic array power loss setting method

Abstract

The invention relates to a method for setting power loss of a mountain photovoltaic array. The method comprises the following steps of: establishing a coupling causal spectrum of multi-source factors including terrain, angle, sunlight, temperature, dust and wiring mode to component output; identifying a dominant loss path based on an identification method of a Bayesian network; outputting a structure weight matrix of the power loss, and identifying an inclination angle gt; a temperature gt; a set of influence factors of the wiring topology pair setting; calculating a power loss response index of each component based on the loss path weight; forming a response index thermodynamic diagram by using a space weight interpolation algorithm, and performing array space sensitive partitioning; the high response area is used as a preferential setting target, and the low response area enters an observation maintenance mechanism; by constructing the response index and the correction response potential function based on multi-factor coupling, the power response potential of the component to the setting operation can be accurately identified, the problems of average adjustment and low-efficiency resource allocation in a traditional setting method are effectively avoided, and it is ensured that setting resources are intensively input into a high-response area.

Description

Technical Field

[0001] The invention relates to a method for setting power loss of a mountain photovoltaic array. Background Art

[0002] While current tuning methods for PV array power loss have some basic adjustment capabilities and empirical optimization methods, they still suffer from numerous deficiencies and systemic drawbacks in practical engineering applications. This is particularly true for mountainous power plants with complex terrain, frequent environmental disturbances, and non-standardized component layouts. Traditional tuning methods are often inadequate, primarily in the following areas:

[0003] First, existing tuning methods generally lack a sophisticated mechanism for identifying spatial heterogeneity. Due to the significant topography of mountainous PV power stations, different regions exhibit significant differences in solar radiation, wind speed and ventilation, dust deposition, and temperature rise and heat dissipation. However, most current tuning methods rely on uniform parameter settings across the entire site or use string-level averages, ignoring module-level power variations and localized environmental conditions. This averaging approach results in a lack of precision in tuning strategies, inefficient use of tuning resources (such as adjustable supports, cleaning equipment, and maintenance personnel), and inability to maximize power recovery. Second, traditional tuning relies primarily on periodic and static rule-based execution, such as scheduled cleaning, fixed angle adjustment cycles, and setpoint tracking. These highly prescriptive and slow-to-respond strategies are ill-suited to addressing the frequent changes in shading, dramatic local microclimate variations, and inconsistent module degradation rates found in mountainous environments. Once the dominant loss factor changes (for example, from dust accumulation to temperature dominance), the system often fails to quickly identify and adjust its strategy, resulting in frequent over- or mis-tuning. This wastes resources and can exacerbate potential risks such as uneven module loading and hot spot effects.

[0004] Most current approaches lack data-driven causal analysis capabilities, relying heavily on operator experience or static rules for strategy selection and lacking dynamic strategy adaptation mechanisms. While some systems have introduced AI components or machine learning prediction modules, most remain at the statistical correlation analysis stage, unable to truly identify the paths of action and changing trends of different loss factors, let alone update strategy selection logic through automatic learning. When multiple factors overlap, the system cannot distinguish between primary and secondary causes, leading to incorrect strategy selection or ineffective execution. Furthermore, these experience-based systems are unable to quickly reuse models or strategies for newly built sites or areas with missing data, resulting in strong site coupling and poor scalability. The granularity and intelligence of the tuning actions themselves are insufficient. Most existing systems use the sole criterion of whether to adjust or not, lacking the ability to make precise decisions about the adjustment range, tuning priorities, and tuning combination strategies. For example, in angle adjustment operations, only consideration is given to whether to adjust, without considering the marginal impact of the specific adjustment on benefits in different regions. In electrical topology tuning, only large-scale series and parallel wiring mode modifications are supported, not dynamic switching at the component level. This limits the optimal balance between benefits and costs in tuning solutions. Furthermore, cleaning strategies are often determined based on whether to clean the entire group, without considering factors such as the location of contamination hotspots or differences in contamination between edge and center components. This leads to excessively high tuning costs and significantly diminishing marginal benefits. Significant shortcomings exist in post-tuning performance evaluation and strategy learning capabilities. Most systems lack a clear mechanism for evaluating tuning benefits. Even after tuning actions are executed, a benefit judgment system based on data feedback is not established, preventing a closed-loop learning path between tuning results and strategy selection. This results in the strategy library remaining static, unupdated after being set, and unable to self-optimize with accumulated operating experience. Furthermore, the lack of a short-term evaluation mechanism prevents timely assessment of the immediate effects of fine-tuning operations. Consequently, tuning strategies often rely on monthly or quarterly power generation statistical analysis, resulting in delayed response and failure to meet the needs of real-time intelligent operation and maintenance. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for adjusting the power loss of a mountain photovoltaic array, thereby solving some of the drawbacks and deficiencies pointed out in the background art.

[0006] The present invention solves the above-mentioned technical problems by adopting the following technical solutions:

[0007] Establish a coupled causal map of the impact of multiple factors, including terrain, angle, sunlight, temperature, dust, and wiring method, on module output. Identify the dominant loss path using a Bayesian network-based identification method. Output the structural weight matrix of power loss and identify the influencing factor group on setting, including inclination > temperature > wiring topology.

[0008] Based on the loss path weights, the power loss response index of each component is calculated. A spatial weight interpolation algorithm is used to generate a response index heat map, which is then used to partition the array into sensitive spatial zones. High-response areas are prioritized for tuning, while low-response areas are subject to observation and maintenance. A simulator is built for each tuning action based on historical data training, constructing a tuning action space. Nonlinear regression or neural networks are then used to evaluate the synergistic benefits of the combined strategies.

[0009] The system continuously tracks the power output changes in high-response areas and their drift trends with meteorological factors in time windows. If the system identifies a loss path offset, it loads the corresponding alternative tuning strategy, implements fine-tuning, and evaluates its short-term benefits. If the results are good, it enters a new round of strategy iteration.

[0010] Furthermore, the method for performing array space sensitive partitioning includes the following steps:

[0011] S1. Obtaining operating parameter data of each photovoltaic module in a mountain photovoltaic array, including power output, shading degree, inclination angle, terrain coordinates, solar irradiation and temperature information;

[0012] S2. Calculate the power variation of each component under the historical tuning behavior and construct a component response index based on the historical power variation to reflect the actual power response capability to the tuning measures.

[0013] S3. Use spatial interpolation algorithm to perform spatial continuous modeling on discrete response indices to form array response index heat map;

[0014] S4. Divide the entire photovoltaic array into high-response, medium-response, and low-response areas; identify the high-response area as a priority tuning target, and perform tuning operations including angle adjustment, component cleaning, and electrical connection optimization; implement an observation and maintenance strategy for the low-response area, and do not implement resource-based tuning operations.

[0015] Furthermore, during the response index construction process, the spatial location of the component and the shading pattern are combined to establish a causal impact path to identify the influencing factors of power loss.

[0016] Furthermore, the spatial interpolation algorithm includes kriging interpolation, inverse distance weighted interpolation or spline surface fitting method to generate a continuous response index distribution map; the identification of the high response area is based on the response index threshold setting, and the priority division supports correction based on the terrain contour distribution to adapt to the complex landform characteristics of the mountains.

[0017] Specifically, the above solution collects real-time operating information such as the output power, current, irradiation, module temperature and position coordinates of each photovoltaic module. Based on the module's response to past tuning behavior, its response index R is calculated. i, as a quantitative expression of the effect of the setting behavior. Use spatial interpolation algorithms such as Kriging, Inverse Distance Weighting (IDW) or spline fitting to convert the discrete points R i Exponentially transforms to a continuous response distribution plot R(x,y).

[0018] In order to improve the response prediction accuracy of complex mountainous areas, the following self-created response potential function model is proposed:

[0019]

[0020] in:

[0021] Ψ(x,y): represents the modified response potential function value at position $(x,y)$, which serves as the final setting priority evaluation indicator; R(x,y): is the initial response index value after spatial interpolation; θ(x,y): is the angle function between the component orientation and the solar incidence angle; It indicates the intensity of the terrain gradient at that point, that is, the slope (undulation intensity) of the contour line at that point; α, β: are empirical correction coefficients used to adjust the proportion of the influence of angle deviation and terrain interference on the setting priority.

[0022] According to the distribution characteristics of Ψ(x,y), set the threshold Ψ th The PV array area is classified and partitioned, with areas above a threshold marked as high-response zones for priority tuning. Furthermore, terrain corrections are introduced for areas with steep slopes to prevent strategy failure. Refined tuning is performed based on the results in high-response zones, and the tuning results are evaluated using short-term operating data, which is then fed back into the response model for adaptive correction and optimization.

[0023] Furthermore, the priority execution order of the setting operations is determined by deduction of a power gain prediction model, which is used to evaluate the synergistic gain effects of different setting actions on each response area.

[0024] Furthermore, the method for implementing fine-tuning and evaluating its short-term benefits includes:

[0025] S1. Real-time collection of operating parameters of each component in the mountain photovoltaic array, including power output, current, voltage, component temperature, light intensity, shading status and external environmental factors;

[0026] S2. Constructing a multi-period performance trend relationship based on the operating parameters to monitor the dominant influencing factors of power loss; when it is identified that the dominant factor of the current power loss path has shifted compared to the historical operating state, it is determined that the loss path is shifting;

[0027] S3. Load the corresponding alternative tuning strategy template for the identified dominant factors, including tilt adjustment, component cleaning, wiring structure adjustment, or regional power scheduling; and perform small-scale tuning operations on the identified areas.

[0028] S4. Evaluate the power gain, current consistency, system load balance and other indicators of the tuning area within a short period after fine-tuning to construct a short-term tuning benefit vector. If the evaluation results show that the tuning strategy is effective, the strategy is incorporated into the subsequent tuning main process. If it is not effective, it is recorded in the strategy correction library for subsequent relationship optimization.

[0029] Furthermore, the alternative setting strategy template pre-establishes a matching strategy library for different main causes of loss, and schedules the call based on the identification results before setting; the fine-tuning operation includes adjustment of the component inclination within the range of ±5°, local switching of the string wiring method, or short-term fixed-point cleaning operation of the target component.

[0030] Furthermore, the short-term tuning benefit evaluation is comprehensively judged based on the power increase after tuning, the component output consistency improvement rate and the system operation smoothness; when the tuning strategy is verified to be effective and repeated, its output weight will be increased and it will be given priority in being included in the recommended strategy set of the subsequent tuning process.

[0031] In the above technical solution, after performing tuning operations including tilt adjustment, cleaning, and wiring optimization, operating data is collected from the target area within a short period (5 to 60 minutes), including actual power changes, output consistency of each component, and current fluctuation. The three indicators of power increase, output consistency improvement rate, and system operation stability are jointly expressed to construct the following tuning benefit function:

[0032]

[0033] in:

[0034] Λ t : The short-term comprehensive benefit value of the tuning strategy at time t; ΔP t : Average power increase of unit components in the setting area; Σ t : The improvement rate of component output consistency after adjustment (such as the standard deviation reduction ratio); Ω t : The system current fluctuation rate after setting, which represents the inverse index of system operation stability; γ, η: empirical adjustment factors, which respectively represent the weight ratio of consistency and fluctuation in the overall evaluation.

[0035] The power loss calibration method for mountain photovoltaic arrays in this invention addresses issues such as non-uniform shading, variable environments, and electrical structure mismatch that exist in photovoltaic systems operating in complex mountainous terrain. It proposes a full-process calibration method from data-driven, causal analysis, response modeling, to strategy execution and feedback. This method has the following significant benefits:

[0036] By constructing a response index and modified response potential function based on multi-factor coupling, the power response potential of the components to the setting operation can be accurately identified, effectively avoiding the problems of average adjustment and inefficient resource allocation in traditional setting methods, ensuring that setting resources are concentrated in high-response areas, and achieving maximum power generation gain with minimal intervention.

[0037] Using multi-period performance trend analysis and causal network modeling, the system can dynamically identify the dominant factors driving power loss (e.g., from shading to temperature) across different time periods and regions, triggering corresponding tuning strategy templates. This allows for an upgrade from static tuning to dynamic perception and response, particularly adapting to the frequent changes in climate, shading, and temperature in mountainous environments. By establishing a library of primary cause-strategy matching templates and integrating them with machine learning models (such as gradient boosting classification and neural network gain prediction models), the system automatically recommends optimal tuning strategies based on historical experience. This system also provides short-term verification and feedback based on actual post-tuning gains, enabling self-learning, self-correction, and self-evolution of strategies, reducing reliance on human experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of the method for adjusting the power loss of a mountain photovoltaic array according to the present invention.

[0039] Figure 2 This is a flow chart of the array space sensitive partitioning method of the present invention.

[0040] Figure 3 Flowchart of a method for implementing fine-tuning of the present invention and evaluating its short-term benefits. DETAILED DESCRIPTION

[0041] The following is a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings.

[0042] From a data-driven perspective, a systematic causal analysis mechanism is constructed to identify key factors influencing tuning effectiveness and, based on these factors, formulate precise and efficient tuning strategies. To achieve this goal, a coupled causal graph encompassing multiple sources of physical and environmental factors must be established. This graph, with module output power as the result node, includes terrain parameters (such as slope, aspect, and relative height difference), module mounting angles (such as inclination and azimuth), solar intensity (based on historical and real-time irradiation data), module surface temperature (obtained from temperature sensors or infrared imagery), dust index (estimated from reflectivity changes or manual cleaning records), and wiring configuration (series-parallel topology and inverter connection). Data acquisition relies on multiple sources, including drone aerial surveys and terrain modeling to extract geographic parameters, PV data collectors (such as data loggers, temperature sensors, and irradiance sensors) to collect module-level operating status, and SCADA systems to record system-level operating parameters. External records such as manual maintenance logs and cleaning frequencies are supplemented to construct a complete dataset.

[0043] During the data preprocessing phase, data formats from different sources and time resolutions must be unified. For example, minute-by-minute power data can be time-windowed with hourly cleaning records, and a unified time step can be constructed using a sliding window or weighted average method. To address missing values, interpolation is used to fill gaps in continuous data, while the mean or nearest neighbor is used to fill missing discrete attributes. Continuous variables are also normalized to eliminate dimensionality effects. For abnormal data processing, extreme points are filtered out using the IQR method or Z-score to reduce the interference of faults or sensor errors on model structure learning. In this example, module power output is used as the target variable, and other input variables are candidate causal variables. The key steps in model construction are structure learning and parameter learning. Structure learning uses a greedy search algorithm (such as the K2 algorithm or a scoring-based structure search method) to explore the optimal network topology. Expert knowledge can also be incorporated to impose a priori constraints on some edges (for example, terrain and angle irreversibly affect sunlight, rather than being affected by it). Parameter learning uses maximum likelihood estimation or Bayesian estimation to estimate marginal probabilities and conditional probability distributions from historical samples.

[0044] After the structural learning is completed, the edge weights or conditional probabilities of each edge in the Bayesian network can be quantified as the causal strength between the influencing factors, and further converted into a power loss structure weight matrix. This matrix describes the relative strength of the impact of each input factor on the output power. For example, when the model inference shows that the average power change caused by the change in inclination angle is 12%, the corresponding impact of the temperature change is 7%, and the impact of the change in the wiring method is 4%, a clearly ranked group of influencing factors can be constructed: inclination angle>temperature>wiring topology. This factor group will directly guide the priority setting of the tuning strategy, which means that when resources are limited or tuning actions are limited, priority should be given to adjusting the inclination parameters, followed by improving component ventilation and heat dissipation to reduce temperature impact, and finally optimizing the electrical connection method to improve system matching.

[0045] The power loss response index calculation stage for each component. The response index is used to characterize the sensitivity of the component to the power recovery effect brought about by the tuning measures (such as tilt adjustment, cleaning, electrical wiring optimization, etc.). Its calculation is based on the structural weight matrix, combined with the current geographical spatial location of each component, historical operating status and tuning behavior response history, to construct a weighted combination function, and couple the edge weight of the dominant factor with the component status to output a single value as the response index of the component. Specific data sources include component-level power change data, tuning operation records (execution time, type, amplitude), surrounding environmental parameters, etc. These data are accumulated over a long period of time through data acquisition terminals, environmental sensors and maintenance record systems. The calculation process of the response index requires data standardization and time synchronization to ensure that the index can be compared horizontally across regions and time periods.

[0046] After calculating the response index of the component, it needs to be expanded from discrete points to a continuous spatial level to form a complete array response heat map. This process is achieved through a spatial weighted interpolation algorithm. Commonly used methods include Kriging interpolation, inverse distance weighted interpolation (IDW), and spline surface fitting. In implementation, the inverse distance weighted interpolation algorithm is preferred as the basic model. Its advantage is that it can quickly capture spatial gradient changes and is suitable for mountainous terrain where the response value fluctuates significantly within a short distance. The interpolation input is the component center coordinates (latitude and longitude or local coordinates) and the response index value, and the output is a continuous distribution map of the response index covering the entire photovoltaic field area. This heat map can clearly show the areas with high and low response intensity distribution, providing a basis for spatial sensitive zoning. Based on the response index heat map, the system performs array spatial sensitive zoning operations, and divides the array into high response areas, medium response areas, and low response areas by setting response thresholds. To adapt to the complex mountainous terrain, a digital elevation model (DEM) was introduced based on the initial threshold zoning results to correct for terrain gradients. For example, even steep slopes with high response values ​​may be re-labeled as low-priority areas due to high maintenance costs or low structural stability. The final zoning results drive the tuning resource allocation logic: high-response areas are designated as priority tuning targets, meaning they will be scheduled for tuning in the next cycle; medium-response areas are dynamically rotated as candidate areas; and low-response areas are marked as observation and maintenance areas. The system only performs status monitoring and does not actively perform tuning, thus conserving maintenance resources.

[0047] After completing regional prioritization, to achieve accurate tuning, the system needs to establish a response simulator model for each tuning action. This model predicts the potential power increase for the target component or region under different tuning types and magnitudes. Data sources include historical tuning actions and their corresponding short-term power generation performance changes, forming historical action-response pairs. During the data preprocessing phase, anomalous samples caused by extreme weather or hardware failures must be removed, and operating parameters (such as angle change and cleaning frequency) must be normalized to improve the generalization capability of model training. For each tuning type (such as tilt adjustment), a separate nonlinear regression model, such as support vector regression (SVR) or random forest regressor, can be trained. For strategic collaborative simulation of tuning action combinations, a feedforward neural network is introduced. Its input is the tuning action vector (a multidimensional tuning parameter combination), and its output is the predicted combined benefit score. The network structure is typically a three-layer fully connected neural network. The number of hidden layer nodes depends on the input dimensionality and sample size. The ReLU activation function is generally used, the mean squared error is the loss function, and the Adam optimizer is used to improve convergence speed.

[0048] The training process is based on historical sample sets, and cross-validation is used to ensure the generalization ability of the model. After training is completed, the simulator model can accept the current partition status and the combination of candidate tuning parameters as input during the actual deployment phase, quickly evaluate each tuning action or combination strategy, and output a benefit score. The system combines all candidate combinations into a tuning action space, uses this space as the optimization object, and uses a strategy filter (which can introduce greedy search, genetic algorithms, or reinforcement learning) to find the optimal tuning solution under the current constraints. After the optimal strategy is fed back to the execution module, it is actually deployed in the priority area, and the actual power generation change data is again included in the historical sample pool to achieve model self-evolution.

[0049] In order to achieve dynamic tuning optimization and adapt to the non-static characteristics of environmental disturbances and system changes, the system needs to perform continuous tracking and feedback adjustments in units of time windows, especially high-frequency dynamic monitoring of previously identified high-response areas. Specifically, after the system demarcates the high-response area, it uses a sliding time window mechanism (such as granularity of 5 minutes, 15 minutes, 1 hour, etc.) to collect the power output data of each component in the area in real time, and correlate it with the meteorological factors of the corresponding time period for tracking, including but not limited to irradiation intensity, wind speed, air temperature, humidity, shading changes (which can be obtained by photovoltaic image recognition or light deviation monitoring), etc. Data collection relies on high-resolution data collectors deployed on-site (collecting voltage, current, temperature, light and other indicators) and connected meteorological stations or micro-meteorological node data, and is uniformly uploaded, cleaned and managed through the SCADA platform.

[0050] Time alignment ensures that each power data entry is organized with the same time step as the meteorological data. Outliers are filtered, such as data jumps caused by irradiance being zero but power exceeding a threshold, or inverter failures. Statistical analysis (such as median deviation) is used to eliminate unreliable data. Normalization is then performed to eliminate training bias introduced by varying data dimensions. Furthermore, continuous data is smoothed using low-pass filtering or sliding averages to reduce noise interference and enhance trend identification sensitivity. Based on the tracking data, the system constructs a time series drift analysis model to determine whether the dominant factor in the current power loss path has shifted. For example, if the original power decline trend was highly correlated with shading changes, but the shading remains stable while the temperature rises within the current sliding window, causing a continuous decline in output, this indicates a shift in the dominant loss path. This determination relies on correlation analysis, sliding window mutual information calculation, or a micro-LSTM prediction model to assess the partial derivative contributions of different factors to power changes. The model structure can be a multi-input, single-output neural network, where the input is the series of meteorological factor changes within the time window and the output is the power change prediction residual. This model can be used to infer the degree to which different factors control the current loss trend.

[0051] Once a loss path shift is identified, the system loads an alternative tuning strategy that matches the new dominant factor. These strategy templates have been mapped through historical tuning data training. For example, when high temperature dominates, strategies such as increasing ventilation spacing and adjusting bracket angles to reduce heat absorption area are likely to be introduced. When dust dominates, a cleaning-first strategy is loaded. The tuning system will call up executable fine-tuning operations from the strategy library based on the current component location, the type of controllable equipment, and resource scheduling capabilities. Fine-tuning operations are characterized by small amplitude and low interference, such as fine-tuning the tilt angle by ±2°, adjusting local MPPT parameters, and performing short-term local cleaning of designated components. After the fine-tuning operation is implemented, the system immediately enters the short-term evaluation phase.

[0052] During the evaluation phase, the power change data before and after fine-tuning in the sliding window, the change in the power output standard deviation (to judge consistency), and the change in inverter efficiency (to judge conversion capability) are used to construct a set of adjustment benefit evaluation indicators. The core evaluation indicators are comprehensive short-term benefit functions, such as: power improvement amplitude, output consistency improvement rate, degree of reduction in operating fluctuations, etc., which are usually constructed as a composite weighted function, and the weights are given by the historical strategy effect regression model. If the evaluation function value exceeds the historical benchmark (or model threshold), that is, the fine-tuning is judged to be an effective strategy, then the strategy is included in the effective strategy set under the current environmental state, its adjustment weight will be increased, and it will be recorded in the strategy evolution library for subsequent training data of the intelligent recommendation system; if the evaluation does not meet the standards, the strategy will be marked as low priority or temporarily frozen, and feedback will be sent to the optimization module to correct its usage scenario description or parameter boundaries.

[0053] The tuning system will then update the current strategy recommendation ranking based on the strategy benefit results, prioritizing recently validated strategies in the next round of tuning and scheduling, forming a closed loop of trial-and-error, validation-and-evolution strategies. In subsequent strategy recommendation models, the system can employ reinforcement learning frameworks (such as DQN) or policy gradient methods to adjust behavior output probabilities based on historical strategy scores, further enabling self-learning and long-term evolution of tuning actions, giving the entire system intelligent tuning capabilities to adapt to complex and changing mountainous environments. This approach transcends the limitations of traditional periodic fixed strategies and subjective human judgment, truly realizing a real-time data-driven, feedback-closed-loop photovoltaic tuning optimization mechanism.

[0054] Example 1:

[0055] Mr. Li, a new energy intelligent operations engineer, was overseeing the tuning and optimization of the photovoltaic array at a typical mountain power station in Yunnan, China. The station, with an installed capacity of 12 MWp, was located on several terraced slopes with an elevation difference of 65 meters. The arrays were unevenly distributed, and some areas were experiencing issues such as tree obstruction, dust accumulation, and inverter mismatch. He recently discovered that total power generation was approximately 18% below design. Mr. Li decided to employ a new generation of spatially sensitive array zoning methods for intelligent tuning of the system to restore power generation efficiency.

[0056] The first step is S1: Obtaining operating parameter data. Engineer Li used the SCADA system operating data of the power station for the past 30 days and combined it with the high-precision DEM modeling results of the drone to extract the data of each component, including its position coordinates (x, y, z), inclination angle (e.g., 32° to 38°), real-time power output (5-minute granularity), incident light irradiance (average value between 720 and 960 W / m 2 The data included over 110,000 records, which, after time synchronization, missing value interpolation, and standardization, served as the foundation for the tuning response model.

[0057] Then enter S2: Construct response index. Li extracted the past tuning behavior data (such as angle adjustment, cleaning frequency, MPPT parameter modification, etc.) and the power increase of the component before and after each tuning behavior, and calculated the average improvement rate under the influence of unit tuning. For example, a component number C2317 in the southeast area brought 6%, 5%, and 8% power increases in three angle adjustments respectively. Its response index is set to After executing this process for all components, a set of discrete response index data is obtained, ranging from approximately 1.5 to 10.8, with an average of 4.7. These indices will serve as the raw point data for subsequent spatial interpolation.

[0058] Then execute S3: spatial interpolation modeling. Mr. Li chose inverse distance weighted interpolation (IDW), and used the component coordinates and response index values ​​as input variables to generate a continuous response index distribution map (heat map) covering the entire power station area. In order to enhance the smoothness of edge transition, the power parameter p = 2.5 in IDW is used, which means that closer points contribute more to the interpolation results. During the interpolation process, he found that the southeast step area and the northwest slope area showed obvious high response index clusters (greater than 7.0), while the response index near the edge of the blocking woods and the central inverter area was generally low (less than 3.5). After the interpolation is completed, a high-resolution response heat map is obtained. The darker the area on the map, the higher the setting potential. The heat map data supports subsequent strategy deployment.

[0059] Finally, S4: Area Division and Adjustment Execution is carried out. Mr. Li divides the area according to the response index threshold: i ≥7.0 is defined as high response area, 4.5 <R i <7.0 is defined as the medium response area, R i≤4.5 is defined as a low response area. After overlaying the partition boundaries on the thermal map, it was found that the high response area accounted for 28% of the array area, mainly concentrated in the southeast slope angle inconsistent area and the back slope offset installation area. Subsequently, Engineer Li deployed an angle adjustment robot to implement ±3° bracket adjustment in this area, and arranged a cleaning robot to perform local cleaning on modules with heavy surface contamination. At the same time, for several component groups with high response but large current fluctuations, the group string wiring method was changed from 6 strings and 1 parallel to 4 strings and 1 parallel to match the inverter input voltage curve. The medium response area was included in the observation list for the next stage of evaluation; the low response area (such as the northwest corner edge and the tree shade interference area) was set as a monitoring object only, and no active adjustment was arranged to reduce maintenance expenses.

[0060] Within 72 hours of implementation, the system conducted a comparative analysis of power output in the high-response area and found that the average module power in the tuning area increased by 9.4%, exceeding the 0.8% increase due to natural fluctuations in the medium-response area. Output stability also improved significantly, with the standard deviation of inter-module power decreasing by 22% and system current matching improving by 13%. Based on the tuning system's feedback mechanism, this round of tuning strategy was marked as highly effective, with its score increasing from 0.61 to 0.87. This strategy was then recorded in the strategy recommendation system for future scenario matching.

[0061] To further improve the accuracy of the strategy, Mr. Li decided to introduce the causal path analysis method into the response index construction proposed in this embodiment. By combining the spatial location and shading pattern of the components, he deeply identified the key influencing factors that actually dominate the power loss and updated the construction logic of the response index.

[0062] The first step of this method is to establish the relationship between the spatial position of the component and the shading structure. Mr. Li retrieved the high-precision digital elevation model (DEM) generated by drone aerial survey, combined with the GIS geographic information system and photovoltaic layout map, and accurately calibrated the relative elevation, slope (such as 15°, 22°, etc.), azimuth (such as 10° south-west) of each component and its geographical location in the array. He paid special attention to the southeast step area and both sides of the central channel, where the terrain changes dramatically and the shading pattern is complex. By importing the local sun trajectory and mountain shadow projection data for the past 90 days through sunshine simulation software (such as SolarAnywhere), Mr. Li established a probability function for each component to be blocked during the daily sunshine time, and output the proportion of shading coverage time. For example, the shading time of component C1045 accounts for 31% of the total illumination time, while the component C3072 is only 4%.

[0063] To establish a causal relationship between these spatial factors and the response index, Li used a causal inference model for data-driven analysis. He chose a structured Bayesian network as the causal path learning model, with input variables including: component height difference Δz, slope θ, orientation α, occlusion coverage ratio s, historical cleaning frequency f, angle adjustment times a, and mean temperature T. avg , and the response variable is the average power increase amplitude ΔP after the previous tuning. By constructing the network structure through the K2 greedy search algorithm, Li found that the shading ratio s and the slope θ constitute the main path affecting the power increase, and the path confidence is as high as 0.86. Although there is a correlation between the number of angle adjustments a and ΔP, it is only significant when the slope angle θ is greater than 20°, indicating that the angle adjustment has a significant effect on high-slope components and is weaker in low-slope areas. In addition, the temperature T avg The path strength with ΔP is 0.44, and the cleaning frequency f with ΔP is 0.38, indicating that temperature has the second greatest influence, followed by dust.

[0064] Based on this causal structure diagram, Engineer Li adjusted the way the response index is constructed. Originally, the response index was calculated based solely on the average of the power increase before and after tuning (for example, the response index of component C1101 is 6.7). Now, the weights of key variables in the causal path are adjusted. Take the following logic as an example:

[0065] R′ i =ΔP i ·(1+ω s ·s i +ω θ ·θ i )

[0066] where s i is the component occlusion ratio, θ i is the slope angle, and the weights are determined by the normalization of the edge strength in the Bayesian network, which are ω s =0.5,ω θ =0.3. Substitute the data of component C1101: ΔP=6.7%, s i =0.24,θ i =21°, then the corrected response index is:

[0067]

[0068] For another component C3072, ΔP is 6.2%, but s i Only 0.04, θ i The corrected response index is:

[0069]

[0070] This means that while the two components showed similar improvements before and after tuning, their response indices differed significantly after the causal correction. The system prioritized the area containing C1101 for subsequent tuning, while relegating C3072 to a suboptimal level. This causal weight correction mechanism effectively addressed tuning bias caused by strong surface response but unresolved root causes of obstruction.

[0071] After applying the updated response index model, Engineer Li regenerated the array heat map and adjusted the spatially sensitive partition boundaries. He discovered that some previously moderately responsive areas had been reclassified as high-responsive areas due to severe terrain obstruction and steep inclination angles. The strategic scheduling system was updated accordingly. The subsequent tuning results significantly improved. After adjusting the bracket angles and clearing obstructions in the reclassified high-responsive areas, the average module power increased by 12.3% over 24 hours, significantly exceeding the original model's estimated 7.5%. The system ultimately assessed that this causal weighted modeling mechanism improved the efficiency of the tuning strategy by approximately 65%.

[0072] The zoning method proposed in this embodiment, based on spatial interpolation of response indices and a modified response potential function model, accurately identifies areas with the greatest potential for tuning benefits by constructing a continuous spatial distribution function Ψ(x,y) that reflects the tuning potential of components. This method has significant advantages, especially in mountainous environments with complex terrain.

[0073] First, in this round of optimization, Li collected real-time operating data for all array components, including the output power (unit W), current (unit A), irradiation intensity (unit W / m 2 ), component temperature (unit: °C), and installation location coordinates (latitude, longitude, and altitude). The power station has a total of 13,200 components, each of which records data in a 10-minute cycle, resulting in up to 2 million historical operation records. To extract the tuning potential, Engineer Li calculated the response index R for each component. i , which is the average unity tuning gain of the component after historical tuning actions (such as tilt adjustment, local cleaning, and electrical reorganization), is defined as follows:

[0074]

[0075] where ΔP i,j is the power increment brought by the jth setting (unit: W), A i,j is the corresponding setting amplitude (such as angle change or cleaning area), N i The number of historical adjustments. 0537 Take the response data as an example: three adjustments bring about 15W, 12W, and 14W increases respectively, and the adjustment ranges are 2.5°, 1.8°, and 2.2°.

[0076]

[0077] He then added all the components' R i The values ​​are combined with their (x, y) coordinates, and the inverse distance weighted interpolation (IDW) algorithm is used to generate a preliminary response index heat map R(x, y). This method sets the power exponent p = 2.2 and performs a weighted average of the response values ​​of each point (x, y), with closer components having higher weights. During the interpolation process, Li found that the response index of components located on the south slope was generally higher (6.5 to 8.9), while some severely obstructed areas on the north slope were below 3.0. To enhance image smoothness, he compared spline fitting with IDW and ultimately chose IDW as the main algorithm because it handles slope boundaries more sharply.

[0078] After obtaining the continuous graph of R(x,y), Mr. Li introduced the modified response potential function model of this embodiment:

[0079]

[0080] Where θ(x,y) is the angle between the module's orientation and the solar incidence angle at noon on that day (in radians), which is obtained by matching the geographical orientation with the solar position data. If the module's orientation is 10° south-west and the sun's position is due south, the angle is θ = 10° = 0.1745 radians. Terrain gradient For the change in contour slope, the local slope is calculated using DEM data and the unit is percentage (e.g. 26% is an area with more severe fluctuations).

[0081] Li set the correction coefficient for this model: α = 0.4, which is used to amplify the penalty effect of angle deviation; β = 0.6, which is used to adjust the penalty weight of strategy failure caused by severe terrain fluctuations. These two values ​​are determined by regression fitting based on historical strategy effects. The recommended range is α∈[0.2,0.6], β∈[0.4,0.8]. 0537 Taking the point as an example, its R(x,y)=6.34,θ(x,y)=0.1745, (i.e. 23% slope), then:

[0082]

[0083] Li Gong generates a modified response potential heat map based on the full-field Ψ(x,y) value, and calculates the value according to the threshold Ψ th =5.5 for regional division. Values ​​above this value are high-response areas, which are used for key deployment in this round and cover an area of ​​about 31%; 4.0-5.5 are medium-response areas, which enter the strategy rotation pool; and values ​​below 4.0 are low-response areas, which are included in the passive observation and remote monitoring group and do not require manual or equipment intervention.

[0084] Finally, Engineer Li performed fine tuning in the high-response area: adjusting the module tilt by ±2°, activating a cleaning robot to perform a fine 22cm cleaning path in locally contaminated areas, optimizing wiring in areas with significant line losses, and adjusting the slope parameters of the inverter's maximum power point tracking (MPPT) curve. Twelve hours after the tuning, real-time monitoring data showed an 11.2% increase in average power in the high-response area, exceeding the 3.1% increase in the medium-response area. The standard deviation of output power decreased by 19%, and overall power generation efficiency increased by 8.7% year-on-year. The tuning feedback data was automatically incorporated into the response index model, and the system entered a new round of adaptive corrections.

[0085] Due to the large area of ​​the high-response area and limited setting resources (such as cleaning robots, adjustable angle bracket controllers, operation and maintenance time windows, etc.), Engineer Li used the power gain prediction model deduction mechanism proposed in this embodiment to optimize the setting sequence. By simulating the power generation benefits of different setting combination strategies in each response sub-area, he found the setting priority sequence with the maximum gain and the best synergy, and realized intelligent control.

[0086] First, Mr. Li divided the high response area into 6 sub-areas (Z1~Z6), each of which is about 150~300 square meters in area and consists of multiple components. In the Z1 area, the average corrected response potential value of the components is 6.23; in the Z2 area, it is 5.88, in Z3, it is 6.41, in Z4, it is 6.12, in Z5, it is 5.95, and in Z6, it is 6.05. The setting actions that can be performed in each area include: A1-tilt adjustment (±2~3°), A2-surface cleaning, A3-wiring structure optimization, and A4-inverter parameter adjustment. Since these setting actions may be superimposed in certain areas, Mr. Li needs to determine whether to do A1 or A2 first, or whether A3 can be performed only on part of the area to achieve the optimal gain. To this end, Mr. Li used the power gain prediction model proposed in this embodiment. The model structure is a multi-input single-output feedforward neural network (FNN). The input vectors include: the corrected response potential value Ψ of the current area, the average irradiation value of the components, the power difference before and after the most recent setting, the type and amplitude of the planned setting action (such as tilt adjustment 2.5°, cleaning level = 2, etc.). The output is the unit power increase (unit W / m2) expected to be brought about by the setting action combination in the area. 2 The network has three layers. The first layer is the input layer (7 neurons, corresponding to the 7 input features mentioned above). The two hidden layers have 16 and 8 neurons respectively. The activation function is ReLU. The output layer is a linear regression node. The training loss function is MSE (mean squared error). The optimizer is Adam. The training set uses a total of 80,000 tuning actions and effect data from the power station over the past two years.

[0087] Taking the Z3 area as an example, Mr. Li tested the following four setting action combination strategies:

[0088] Solution 1: Only perform A1 (tilt adjustment 2°)

[0089] Solution 2: Perform A1+A2 (tilt adjustment 2°+cleaning)

[0090] Option 3: Perform A2+A3 (cleaning+rewiring)

[0091] Solution 4: Execute A1+A3+A4 (full action set)

[0092] Substituting the input features of the Z3 region into the model: Ψ = 6.41, the average irradiance value is 890 W / m 2 , the power increase before and after tuning was 6% in the past, and the predicted gain values ​​of each combination were simulated as follows:

[0093] Option 1: Expected increase of 12.3W / m 2

[0094] Option 2: Expected increase of 17.6W / m 2

[0095] Option 3: Expected increase of 14.2W / m 2

[0096] Option 4: Expected increase of 18.1W / m 2

[0097] Although Option 4 offers slightly higher returns, it involves full tuning, consuming significant equipment resources, and taking a long time to operate. The system's built-in resource weighting model calculates the required operational load index (expressed as a combination of equipment hours, control system load, and dispatching costs) for each unit of power increase as follows:

[0098] Option 1: 1.8

[0099] Option 2: 2.9

[0100] Option 3: 3.1

[0101] Option 4: 4.5

[0102] Dividing the gain value by the operating load index gives the cost-performance index:

[0103] Option 1: 6.83

[0104] Option 2: 6.07

[0105] Option 3: 4.58

[0106] Solution 4: 4.02

[0107] Finally, the system decided to prioritize the implementation of Plan 1 and fine-tune the angle of the Z3 area. Because the response potential of the Z1 area is lower than that of Z3 but slightly higher than that of Z5, the system selected Plan 2 for it, a combination strategy of cleaning + angle adjustment. Due to the high slope and complex cable laying in the Z2 and Z5 areas, it was chosen to only adjust the MPPT parameters, that is, to perform the A4 operation. The Z6 area executed A3 (partial series-parallel replacement) because of its complex wiring topology and good performance in the early setting response. After completing the strategy selection, Mr. Li performed the operations in sequence. After 24 hours, the system recorded the power increase data, and the Z3 area increased to 12.9W / m 2 , slightly higher than the model prediction value of 12.3, verifying the model's prediction accuracy. At the same time, the system automatically appends the actual tuning data to the model training data set for subsequent model retraining to improve long-term learning capabilities.

[0108] Example 2:

[0109] Based on Example 1, a mountain photovoltaic power station at an altitude of 1,650 meters in Honghe Prefecture, Yunnan Province, has completed response potential function modeling, response zone division, and strategy priority arrangement. During continuous operation, the system detected abnormal fluctuations in component power in area Z5 (northwest slope), with an average decrease of approximately 11.6% compared to a week ago. Engineer Li immediately decided to initiate the fine-tuning and short-term benefit evaluation scheme proposed in this example and implement it according to the technical paths S1 to S4.

[0110] S1: Real-time collection of operating parameters. Engineer Li uses the photovoltaic monitoring system SCADA and string collection terminal to capture real-time data of 142 panels in the Z5 area. The collected indicators include: current power output (unit W), current (unit A), voltage (unit V), panel surface temperature (℃), solar irradiation (W / m 2 ), occlusion index (based on the occlusion image recognition score, between 0 and 1), wind speed, humidity, and ambient temperature. Taking the C507 module as an example, its current power is 174W (a decrease of about 14% compared to the design value), the current is 4.5A, the module temperature is 68℃, the occlusion index is 0.08, and the light intensity is 910W / m 2 , wind speed 2.1m / s.

[0111] S2: Construct multi-period performance trends and identify the shift in dominant factors. Engineer Li called the component performance data and operation logs of the power station in the past 72 hours, compared them with the operating average of the same period last month, constructed a three-segment time window (short period 3 hours, medium period 24 hours, long period 72 hours) trend model, and adopted a sliding window correlation analysis method. The system determined that the illumination of most components in the Z5 area did not fluctuate significantly (the correlation remained above 0.92), and the shading status was also relatively stable, but the negative correlation between temperature change and power drop increased from -0.45 to -0.78. Based on this, the system identified that the dominant factor of power loss shifted from shading to temperature influence. Based on this, Engineer Li called the preset rules to determine: Since the high temperature in summer is approaching 75°C, some components are close to the performance degradation threshold (about 70°C), and a temperature suppression tuning strategy needs to be loaded.

[0112] S3: Load the alternative tuning strategy and implement fine-tuning. The system automatically retrieves a tuning template for high-temperature-dominated loss paths and recommends implementing a lightweight, ventilation-enhanced bracket angle optimization strategy for the affected area. This involves slightly increasing the angle by 1.5° to increase back airflow. A targeted water curtain spraying experiment is also conducted to test its cooling effect. Engineer Li selected four modules, C507, C510, C514, and C520, in the Z5 area as samples for fine-tuning. The angle was adjusted from 32° to 33.5°, and the spray system was activated, spraying for 1 minute and 10 minutes at a time, for 30 minutes.

[0113] S4: Short-term evaluation of fine-tuning benefits. Within 1 hour after the adjustment, the system continuously collects parameters of components such as C507 at a 5-minute granularity. Evaluation indicators include:

[0114] Power gain (ΔP): increased from 174W to 191W, an increase of 17W, or 9.77%;

[0115] Current consistency (σ I ): The original standard deviation was 0.83A, which was reduced to 0.61A, improving consistency by 26.5%;

[0116] System load balance (LoadIndex): The power fluctuation on the input side of inverter Z5 decreased from ±12.6% to ±6.2%.

[0117] The short-term benefit function set by the system according to this embodiment is:

[0118]

[0119] The weights are set to w1=0.5, w2=0.3, w3=0.2, and the substitution yields:

[0120] Λ=0.5·0.0977+0.3·0.265+0.2·0.5079≈0.04885+0.0795+0.1016≈0.22995

[0121] This result exceeded the locally set fine-tuning strategy benefit threshold of 0.15, prompting the system to determine that the fine-tuning strategy was effective. The system then added the ventilation angle increase and localized spray strategy to the current set of recommended strategies for the Z5 region, raising its priority to Class A. The system also stored the fine-tuning operation and evaluation data in the strategy training log for subsequent regression model optimization. The system also designated this component group as a strategy evolution point, allowing the system to continuously optimize its tuning approach through self-learning in the future.

[0122] If the evaluation result is lower than 0.15 (for example, there is no obvious power recovery or the current consistency decreases), the system will automatically mark the strategy as pending regression verification, reduce its execution probability, and seal it under the same environmental conditions.

[0123] This example establishes a library of tuning strategies that adapt to various dominant loss factors and deploys an intelligent tuning system that can be called and executed in real time. The core of this system is the training process of the strategy-scenario mapping model and the automated execution capability of the actionable module.

[0124] First, Mr. Li organized the operation and maintenance team to analyze the common types of power loss in the operation of the power station in the past two years and divided them into five dominant factors: ① shading dominant type, ② high temperature dominant type, ③ pollution and dust dominant type, ④ electrical mismatch dominant type, and ⑤ seasonal light deviation dominant type. For each type of main cause, the system builds a strategy template library. Each template record contains three types of information: (1) applicable condition feature set; (2) recommended action combination set; (3) execution parameter recommendation. For example, the high temperature dominant template includes ambient temperature>32℃ and component surface temperature>65℃, solar radiation>850W / m 2, shading index <0.15, wind speed <3m / s; recommended actions include increasing the tilt angle by 2-3°, optimizing local ventilation, and fine-tuning the MPPT response curve. After the template was formed, Mr. Li used the 80,000 historical setting and operation data of the power station to match and annotate them as samples for subsequent classification model training. All historical samples were aggregated at hourly granularity to construct a sample feature vector X, including 8-dimensional indicators such as ambient temperature, component temperature, shading index, slope, wind speed, irradiation, historical power deviation rate, and sunshine duration deviation. Label Y is the actual execution strategy type (such as Template_2 represents high-temperature setting). For this multi-class classification problem, Mr. Li used a lightweight gradient boosting tree (LightGBM) to build a classification model, inputting the feature vector and outputting the predicted strategy type number. The model structure consists of 200 trees with a maximum depth of 8. The average accuracy of 10-fold cross-validation reached 92.3%, which is better than the traditional logistic regression model. After the model training is completed, whenever the system identifies a shift in the main cause of power loss (such as a shift from shading to temperature), the model calls the matching strategy template in real time based on the current environmental conditions of the component and generates an action list.

[0125] In actual application, Mr. Li selected the high-temperature setting fine-tuning for the Z3 area in the system. The recommended strategy output by the model is Template_2, and the corresponding specific fine-tuning actions are as follows:

[0126] 1. Module tilt fine-tuning: Within the bracket's adjustable range of ±5°, adjust the angle upward to 33.5° to enhance rear-facing convection. The system reads the recommended angle from the strategy template library and increases it by 2° from the current value. The PLC controls the bracket actuators to send adjustment commands to the corresponding groups, with an adjustment accuracy of ±0.2°.

[0127] 2. Local string wiring switch: The system identified that the original 6-string, 1-parallel connection in this area was causing some weaker current components to reduce the overall string output. Based on the strategy, the system recommended switching to 3-string, 2-parallel on a specific branch to improve overall voltage matching efficiency. This electrical switchover was accomplished through the intelligent circuit breaker system within the programmable PV combiner box, and this operation was performed remotely after authorization by the backend system.

[0128] 3. Fixed-point cleaning operation: Call the pollution tolerance threshold recorded in the strategy library (surface reflectivity drops by more than 15%), and combine the latest drone infrared imaging data to find hot spot clusters C316 and C319. Arrange fixed-point low-pressure water spray robots to go there for cleaning. The spray intensity is 1.2L / min and the duration is 3 minutes. It belongs to the light cleaning level.

[0129] After completing fine-tuning operations, the system records the execution time, operating parameters, execution area, component response, and short-term power gain within one hour of the tuning action, creating a strategy effectiveness log. If a strategy is repeatedly evaluated as highly effective in a given scenario, its strategy weight will be increased and it will be prioritized in future recommended strategy sequences. Conversely, if an action performs poorly in similar scenarios, the system will automatically downgrade its priority and add it to the strategy review queue for manual review and feature adjustment during subsequent model iterations.

[0130] Multi-dimensional operational data is used to make immediate value judgments on tuning behaviors and determine whether to increase their weight in future recommendation strategies. This evaluation was conducted on 116 components in the Z3 area. Engineer Li used 5-minute granularity sampling to continuously collect key operating parameters within a 60-minute window after the tuning was completed, and calculated the short-term benefit value Λ according to the set benefit function. t .

[0131] First, Mr. Li calculated the average power increase ΔP t Before tuning, the average unit module power in the Z3 area was 172.4W. After tuning, the operating data within 15 to 60 minutes showed a steady increase to 186.3W. Therefore:

[0132] ΔP t =186.3-172.4=13.9W

[0133] The second is the output consistency improvement rate Σ t Before tuning, the power standard deviation of the 116 modules in the Z3 area was 21.8W; after tuning, it dropped to 16.5W, and the consistency improvement rate was:

[0134]

[0135] Again, the system current fluctuation rate Ω t Engineer Li obtained the maximum current variation of the Z3 branch through real-time current sampling of the inverter. The standard deviation of the current fluctuation was 2.65A before adjustment, and dropped to 1.93A after adjustment. Therefore:

[0136]

[0137] According to the short-term adjustment benefit function set in this embodiment:

[0138]

[0139] In the system, Mr. Li set the empirical adjustment factors to γ ​​= 0.6 and η = 0.7. This range was determined after a large number of scenario simulations and strategy verifications. The recommended applicable range is γ∈[0.4,0.8] and η∈[0.5,1.0]. Substituting the values ​​into the formula:

[0140]

[0141] The system sets a reference threshold for the effectiveness of the tuning strategy as Λ th =8.0, if it exceeds this value, it is determined to be an efficient strategy. t =10.54>8.0, the system evaluates this combined tilt adjustment + cleaning + wiring structure optimization strategy as effective and marks it as the preferred strategy in the strategy recommendation system. The system then performs two feedback actions: first, it writes this strategy into the strategy recommendation priority table for this site, giving it priority in future deployments under similar environmental conditions (such as high temperatures and large module temperature differences in summer); second, it records this strategy's operation steps, parameters, timeliness, and execution results in the strategy learning library for subsequent deep reinforcement learning model training and strategy library self-evolution.

[0142] In addition, to verify the generalizability of the tuning action, Engineer Li replicated the same strategy in the Z4 region and collected data for comparison. The power increase in the Z4 region was slightly lower, averaging 10.2W, with a consistency improvement rate of 0.19 and a volatility of 0.81. Substituting the same parameters, the following calculations were performed:

[0143]

[0144] If the strategy falls below the set threshold, the system automatically marks it as context-dependent, meaning it's not worth promoting in all regions, and reduces its general recommendation weight. This mechanism ensures that the strategy is only consistently applied in regions with structural profit potential, avoiding a one-size-fits-all approach that wastes resources.

[0145] In summary, this example fully demonstrates the practical feasibility of the short-term adjustment benefit evaluation mechanism proposed in this embodiment. In particular, by collecting power, current and consistency indicators in real time and constructing a composite benefit function, it is possible to accurately identify the value of the strategy. At the same time, the application of dynamic adjustment coefficients enables this method to have scenario-adaptive capabilities. More importantly, the benefit value Λ t The structural form is not only interpretable, but can also be used in subsequent intelligent systems such as reinforcement learning and strategy scoring to form a three-dimensional closed loop of strategy-data-effect. It is one of the key basic modules of future intelligent photovoltaic setting systems.

Claims

1. A method for determining power loss of a mountain photovoltaic array, characterized by: Establish a coupled causal map of the impact of multiple factors, including terrain, angle, sunlight, temperature, dust, and wiring method, on module output. Identify the dominant loss path using a Bayesian network-based identification method. Output the structural weight matrix of power loss and identify the influencing factor group on setting, including inclination > temperature > wiring topology. Based on the loss path weights, the power loss response index of each component is calculated. A spatial weight interpolation algorithm is used to generate a response index heat map, which is then used to partition the array into sensitive spatial zones. High-response areas are prioritized for tuning, while low-response areas are subject to observation and maintenance. A simulator is built for each tuning action based on historical data training, constructing a tuning action space. Nonlinear regression or neural networks are then used to evaluate the synergistic benefits of the combined strategies. Based on the time window, the power output changes in the high response area and its drift trend with meteorological factors are continuously tracked; If the system identifies a loss path deviation, it loads the corresponding alternative tuning strategy, implements fine-tuning, and evaluates its short-term benefits. If the effect is good, it enters a new round of strategy iteration.

2. The method for determining power loss of a mountain photovoltaic array according to claim 1, characterized in that The method for performing array space sensitive partitioning comprises the following steps: S1. Obtaining operating parameter data of each photovoltaic module in a mountain photovoltaic array, including power output, shading degree, inclination angle, terrain coordinates, solar irradiation and temperature information; S2. Calculate the power variation of each component under the historical tuning behavior and construct a component response index based on the historical power variation to reflect the actual power response capability to the tuning measures. S3. Use spatial interpolation algorithm to perform spatial continuous modeling on discrete response indices to form array response index heat map; S4. Divide the entire photovoltaic array into high-response, medium-response, and low-response areas; identify the high-response area as a priority tuning target, and perform tuning operations including angle adjustment, component cleaning, and electrical connection optimization; implement an observation and maintenance strategy for the low-response area, and do not implement resource-based tuning operations.

3. The method for determining power loss of a mountain photovoltaic array according to claim 2, characterized in that During the response index construction process, the spatial location of the component and the shading pattern are combined to establish a causal impact path to identify the influencing factors of power loss.

4. The method for determining power loss of a mountain photovoltaic array according to claim 3, characterized in that The spatial interpolation algorithm includes kriging interpolation, inverse distance weighted interpolation or spline surface fitting method to generate a continuous response index distribution map; the identification of the high response area is based on the response index threshold setting, and the priority division supports correction based on the terrain contour distribution to adapt to the complex landform characteristics of the mountain.

5. The method for determining power loss of a mountain photovoltaic array according to claim 4, characterized in that The priority execution order of the setting operations is determined by deduction of a power gain prediction model, which is used to evaluate the synergistic gain effects of different setting actions on each response area.

6. The method for determining power loss of a mountain photovoltaic array according to claim 1, characterized in that The methods described for implementing fine-tuning and evaluating its short-term benefits include: S1. Real-time collection of operating parameters of each component in the mountain photovoltaic array, including power output, current, voltage, component temperature, light intensity, shading status and external environmental factors; S2. Constructing a multi-period performance trend relationship based on the operating parameters to monitor the dominant influencing factors of power loss; when it is identified that the dominant factor of the current power loss path has shifted compared to the historical operating state, it is determined that the loss path is shifting; S3. Load the corresponding alternative tuning strategy template for the identified dominant factors, including tilt adjustment, component cleaning, wiring structure adjustment, or regional power scheduling; and perform small-scale tuning operations on the identified areas. S4. Evaluate the power gain, current consistency, system load balance and other indicators of the tuning area within a short period after fine-tuning to construct a short-term tuning benefit vector. If the evaluation results show that the tuning strategy is effective, the strategy is incorporated into the subsequent tuning main process. If it is not effective, it is recorded in the strategy correction library for subsequent relationship optimization.

7. The method for determining power loss of a mountain photovoltaic array according to claim 6, characterized in that The alternative setting strategy template pre-establishes a matching strategy library for different main causes of loss, and schedules the call based on the identification results before setting; the fine-tuning operation includes adjusting the component inclination within the range of ±5°, local switching of the string wiring method, or short-term fixed-point cleaning of the target component.

8. The method for determining power loss of a mountain photovoltaic array according to claim 7, characterized in that The short-term tuning benefit evaluation is based on a comprehensive assessment of the power increase after tuning, the component output consistency improvement rate, and the system operation smoothness. When a tuning strategy is verified to be effective and repeated, its output weight will be increased and it will be given priority in being included in the recommended strategy set for subsequent tuning processes.

Images