Method for setting power loss of mountain photovoltaic array

Abstract

The present application relates to the setting method of mountain photovoltaic array power loss, including the establishment of the coupling causal atlas of multiple source factors including terrain, angle, sunshine, temperature, dust, wiring mode to the output of component; the dominant loss path is identified based on the identification method of bayesian network; the structural weight matrix of output power loss is output, the influence factor group including inclination > temperature > wiring topology is identified; based on the weight of loss path, the power loss response index of each component is calculated; the response index thermal map is formed by using the spatial weight interpolation algorithm, and the array space sensitive partition is carried out; the high response area is used as the priority setting target, and the low response area enters the observation maintenance mechanism; the present application can accurately identify the power response potential of the component to the setting operation by constructing the response index and the correction response potential function based on the coupling of multiple factors, effectively avoid the problems of average adjustment and inefficient resource allocation in the traditional setting method, and ensure that the setting resources are concentrated in the high response area.

Description

TECHNICAL FIELD

[0001] The present application relates to a method for setting the power loss of a mountain photovoltaic array. BACKGROUND

[0002] The current setting method for photovoltaic array power loss, although it has basic adjustment ability and empirical optimization means to some extent, still has many deficiencies and systematic disadvantages in practical engineering applications, especially in the face of mountain power stations with complex terrain, frequent environmental disturbances, and non-standardized component arrangement. The traditional setting method is not up to the task, mainly in the following aspects:

[0003] First, the existing setting method generally lacks a fine identification mechanism for spatial heterogeneity. Mountain photovoltaic power stations have significant terrain undulations, and different regions have significant differences in terms of solar radiation, wind speed ventilation, dust deposition, temperature rise, and heat dissipation. However, most current setting methods are still based on uniform parameter setting for the entire field or average values at the string level, ignoring the local nature of component-level power differences and environmental conditions. This average processing method results in a lack of precision in the setting strategy, low efficiency in the use of setting resources (such as adjustable supports, cleaning equipment, and manpower for operation and maintenance), and cannot achieve maximum power recovery. Second, traditional setting mainly relies on periodic and static rule libraries for execution, such as periodic cleaning, fixed angle adjustment period, and set value tracking. This strategy, which is strongly pre-set and slow to react, is difficult to deal with the actual problems of frequent changes in shading, dramatic differences in local microclimate, and inconsistent component degradation rates in mountainous environments. Once the dominant loss factor changes (for example, from dust accumulation to temperature dominance), the system often cannot quickly identify and adjust the strategy, leading to frequent over-adjustment or incorrect adjustment. Not only does this waste resources, but it can also exacerbate potential risks such as uneven component load and hot spot effect.

[0004] Most of the current methods lack data-driven causal analysis capabilities, and are still highly dependent on the experience or static rules of operation and maintenance personnel in strategy selection, lacking dynamic strategy adaptation mechanisms. Although some systems have introduced AI components or machine learning prediction modules, most of them still remain at the stage of statistical correlation analysis, and cannot truly identify the action path and trend of different loss factors, nor can they update the strategy selection logic through automatic learning. When multiple factors overlap, the system cannot determine the primary and secondary causes, leading to strategy selection errors or ineffective execution. In addition, this experience-dependent system cannot quickly reuse models or strategies when facing new stations or data missing areas, and has strong station coupling, poor scalability. The granularity and intelligence level of the setting action itself are insufficient. Most existing systems use adjustment or non-adjustment as the only judgment standard, lacking the ability to make fine decisions on adjustment amplitude, setting priority, and combination strategy. For example, in angle adjustment operations, only whether to adjust is considered, without considering the marginal change of specific adjustment degrees on the revenue of different areas. In electrical topology setting, only large-scale series-parallel connection mode modification is supported, without supporting dynamic switching at the component level, thereby limiting the optimal matching of setting schemes between benefits and costs. At the same time, cleaning strategies often decide whether to clean the entire set, without considering factors such as pollution hot spot location, edge and center component pollution difference, leading to high setting cost and obvious marginal revenue decrement. There are significant shortcomings in the effect evaluation and strategy learning ability after setting execution. Most systems lack a clear setting benefit evaluation mechanism, and even after executing the setting action, they cannot build an efficiency judgment system based on data feedback, thereby failing to form a closed-loop learning path between the setting result and the strategy selection, leading to a static state of the strategy library that is not updated after being set, and the inability to gradually self-optimize with the accumulation of operation experience. In addition, due to the lack of short-period evaluation mechanism, the immediate effect of fine-tuning operations cannot be judged in time, leading to a long-term reliance on monthly or seasonal power generation statistics analysis for setting strategies, with a lagging response that cannot meet the real-time intelligent operation and maintenance requirements. SUMMARY

[0005] The purpose of the present application is to provide a method for setting the power loss of a mountain photovoltaic array, thereby solving some of the problems and deficiencies identified in the background art.

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

[0007] A multi-source factor coupling causal graph is established, including terrain, angle, sunshine, temperature, dust, and wiring mode, to the output of the component; a Bayesian network-based identification method identifies the dominant loss path; a structural weight matrix of output power loss is output, identifying the influence factor group including inclination > temperature > wiring topology for setting;

[0008] Based on the loss path weight, the power loss response index of each component is calculated; the response index thermal map is formed using the spatial weight interpolation algorithm, and the array space sensitive partition is carried out; the high response area is taken as the priority setting target, and the low response area enters the observation maintenance mechanism; based on the historical data, the simulator is trained for each kind of setting action, the setting action space is constructed, and the synergistic benefit of the combined strategy is evaluated by using nonlinear regression or neural network;

[0009] In time window units, the power output change of the high response area and its drift trend with the meteorological factor are continuously tracked; if the system identifies that the loss path deviates, the corresponding alternative setting strategy is loaded; and fine tuning is implemented, and its short-term benefit is evaluated, if the effect is good, a new round of strategy iteration is entered.

[0010] Further, the array space sensitive partition method step comprises:

[0011] S1, obtaining the operating parameter data of each photovoltaic component in the mountain photovoltaic array, including power output, shading degree, inclination, terrain coordinate, solar radiation and temperature information;

[0012] S2, calculating the power change amplitude of each component under the historical setting behavior, and constructing the component response index according to this, which is used to reflect the actual power response ability to the setting measures;

[0013] S3, using a spatial interpolation algorithm to model the discrete response index spatially continuously, forming an array response index thermal map;

[0014] S4, dividing the entire photovoltaic array into high response area, medium response area and low response area; identifying the high response area as the priority setting target, and executing the setting operation including angle adjustment, component cleaning and electrical connection optimization; executing the observation maintenance strategy for the low response area, and not implementing the resource setting operation.

[0015] Further, in the response index construction process, the spatial position and shading mode of the component are combined to establish the causal influence path to identify the influence factor of power loss.

[0016] Further, the spatial interpolation algorithm includes but is not limited to Kriging interpolation method, inverse distance weighted interpolation method 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 the correction based on the terrain contour distribution, to adapt to the complex topographic features of the mountain.

[0017] In the above scheme, specifically: the output power, current, irradiation, component temperature and position coordinate and other operating information of each photovoltaic component are collected in real time. According to the response performance of the component to the past setting behavior, the response index R i, as the quantitative expression of the effect of setting behavior. The spatial interpolation algorithm such as Kriging, inverse distance weighted (IDW) or spline fitting is used to convert the discrete points of R i into a continuous response distribution map R(x, y).

[0018] To improve the response prediction accuracy in complex mountainous areas, the following self-created response potential function model is proposed:

[0019]

[0020] wherein:

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

[0022] According to the distribution characteristics of Ψ(x, y), a threshold value Ψ th Classify and partition the photovoltaic array area, mark the area above the threshold value as a high response area, and use it for priority execution of setting operation; in addition, introduce a terrain factor correction for steep slope areas to avoid strategy failure. According to the high response area result, perform fine tuning, and evaluate the setting effect through short period running data, and feed back to the response model for adaptive correction and optimization.

[0023] Further, the priority execution order of the setting operation is determined by the power gain prediction model, which is used to evaluate the synergistic gain effect of different setting actions on each response area.

[0024] Further, the method of implementing fine tuning and evaluating its short-term benefits comprises:

[0025] S1, real-time acquisition of the running parameters of each component in the mountain photovoltaic array, including power output, current, voltage, component temperature, light intensity, shading state and external environmental factors;

[0026] S2, based on the running parameters, construct a multi-period performance trend relationship to monitor the dominant influence factor of power loss; when the dominant factor migration of the current power loss path is identified compared with the historical running state, it is determined that the loss path deviates;

[0027] S3, for the identified dominant factor, load the corresponding alternative setting strategy template, including inclination adjustment, component cleaning, wiring structure adjustment or regional power scheduling; Implement small range setting fine-tuning operation on the identified area;

[0028] S4, evaluate the power gain, current consistency, system load balance degree and other indicators of the setting area in a short period after fine-tuning, and construct a short-term setting benefit vector; If the evaluation result shows that the setting strategy is effective, the strategy is included in the subsequent setting mainstream process; If it is not effective, it is recorded in the strategy correction library for subsequent relationship optimization.

[0029] Further, the alternative setting strategy template pre-establishes a matching strategy library for different loss main factors, and calls according to the identification result before setting; The fine-tuning operation includes component inclination adjustment within ±5° range, local switching of string wiring mode or short-time point cleaning operation of target component.

[0030] Further, the short-term setting benefit evaluation is based on the comprehensive determination of power improvement amplitude, component output consistency improvement rate and system running smoothness after setting; When the setting strategy is verified to be effective and repeatedly appears, its output weight will be improved, and it will be preferentially included in the recommended strategy set of subsequent setting process.

[0031] In the above technical solution, after performing the setting operation including inclination adjustment, cleaning, wiring optimization, the running data of the target area in a short period (5-60 minutes) is collected, including actual power change, component output consistency, current fluctuation degree. The three indexes of power improvement amplitude, output consistency improvement rate and system running smoothness are jointly represented, and the following setting benefit function is constructed:

[0032]

[0033] Wherein:

[0034] Λ t : Short-term comprehensive benefit value of setting strategy at time t; ΔP t : Average power improvement amplitude of unit component in the setting area; Σ t : Output consistency improvement rate of the component after setting (such as standard deviation reduction ratio); Ω t : System current fluctuation rate after setting, indicating the inverse index of system running stability; γ, η: Empirical adjustment factor, representing the weight proportion of consistency and fluctuation degree in overall evaluation.

[0035] The setting method for power loss of mountain photovoltaic array of the application aims at the problems of non-uniform shading, variable environment, and mismatched electrical structure in the operation of photovoltaic system under complex mountain terrain, and proposes a whole-process setting method from data driving, causal analysis, response modeling to strategy execution and feedback, which has the following remarkable beneficial effects:

[0036] By constructing the response index and the modified response potential function based on multi-factor coupling, the power response potential of the components to the setting operation can be accurately identified, the problems of average adjustment and inefficient resource allocation in the traditional setting method can be effectively avoided, the setting resources can be concentrated in the high-response area, and the maximum power generation gain under the minimum intervention can be realized.

[0037] By using the multi-period performance trend analysis and causal network modeling method, the dominant factors (such as the change from shading to temperature) of power loss in different time periods and different areas can be dynamically identified, so that the corresponding setting strategy template can be triggered, the upgrade from static setting to dynamic perception and response can be realized, and the frequent changes of climate, shading and temperature in mountain environment can be particularly adapted. By establishing the main factor-strategy matching template library and combining the machine learning model (such as the gradient boosting classification model and the neural network gain prediction model), the system can automatically recommend the optimal setting strategy according to the historical experience, and the short-period verification feedback can be performed through the actual income after setting, so that the strategy can be self-learned, self-corrected and self-evolved, and the dependence on artificial experience can be reduced. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The flow chart of the setting method for power loss of mountain photovoltaic array of the application.

[0039] Figure 2 The flow chart of the array space sensitive partitioning method of the application.

[0040] Figure 3 The flow chart of the method for implementing fine tuning and evaluating short-term income of the application. DETAILED DESCRIPTION

[0041] The specific embodiments of the application will be described in detail below with reference to the accompanying drawings.

[0042] A systematic causal analysis mechanism is constructed from a data-driven perspective to identify key factors affecting the setting effect and develop accurate and efficient setting strategies accordingly. To achieve this goal, a coupled causal graph covering multiple source physical and environmental factors is first established, which takes the component output power as the result node and includes terrain parameters (such as slope, slope direction, and relative elevation), component installation angle (such as inclination and azimuth), solar intensity (based on historical and real-time irradiation data), component surface temperature (obtained by temperature sensor or infrared image), dust index (estimated by reflectivity change or manual cleaning record), and wiring mode (series and parallel topology structure and inverter connection) as potential influencing factors. Data acquisition relies on multiple sources, including unmanned aerial vehicle surveying and terrain modeling to extract geographic parameters, photovoltaic data collectors (such as data collectors, temperature sensors, and irradiation sensors) to collect component-level operating status, SCADA systems to record system-level operating parameters, and external records such as manual maintenance logs and cleaning frequency to build a complete data set.

[0043] In the data preprocessing stage, it is necessary to unify the data format of different sources and time resolutions, such as aligning the power data recorded in minutes with the cleaning records at the hour level through time window alignment, constructing uniform time steps through sliding window or weighted average method. For missing value processing, interpolation method is used to fill in the continuity data gaps, and mean or nearest neighbor is used to fill in the discrete attribute missing; at the same time, the continuous variables are standardized to eliminate the influence of dimension. In terms of abnormal data processing, extreme points are filtered out through IQR method or Z-score to reduce the interference of faults or sensor errors on model structure learning. In this embodiment, the component power output is taken as the target variable, and other input variables are candidate causal variables. The key steps of model construction are structure learning and parameter learning. Structure learning uses greedy search algorithm (such as K2 algorithm or scoring-based structure search method) to explore the optimal network topology, and can combine expert knowledge to make prior constraints on some edges (for example, terrain and angle cannot reversely affect solar radiation, but are affected by it). Parameter learning uses maximum likelihood estimation or Bayesian estimation method to estimate the marginal probability and conditional probability distribution from historical samples.

[0044] After structure learning, the edge weights or conditional probabilities of each edge in the Bayesian network can be quantified as the causal strength between the influencing factors, which are further converted into the power loss structure weight matrix. This matrix describes the relative strength of each input factor on the output power. For example, when the model infers that the change in inclination causes an average change in power of 12%, the change in temperature corresponds to an impact of 7%, and the change in wiring method affects 4%, an ordered set of influencing factors can be constructed: inclination > temperature > wiring topology. This factor set will directly guide the priority setting of the tuning strategy, meaning that in the case of limited resources or limited tuning actions, the inclination parameter should be adjusted first, followed by improving component ventilation to reduce temperature effects, and finally optimizing electrical connection to improve system matching.

[0045] Power loss response index calculation phase for each component. The response index is used to characterize the sensitivity of the component to the power recovery effect of the tuning measures (such as inclination adjustment, cleaning, electrical wiring optimization, etc.). Its calculation is based on the structure weight matrix, combined with the current geographical position, historical operating state and tuning behavior response history of each component, to build a weighted combination function. The edge weight of the dominant factor is coupled with the component state to output a single numerical value as the response index of the component. Specific data sources include component-level power change data, tuning operation records (execution time, type, amplitude), and surrounding environmental parameters. These data are accumulated through data collection terminals, environmental sensors, and maintenance record systems over a long period of time. The calculation process of the response index needs to standardize and time-synchronize the data to ensure that the index can be compared horizontally across regions and time periods.

[0046] After calculating the response index of each component, it needs to be extended from discrete points to continuous spatial levels to form a complete array response heat map. This process is achieved through spatial weight interpolation algorithm, commonly used methods include Kriging interpolation, inverse distance weighted interpolation (IDW) and spline surface fitting. In practice, the inverse distance weighted interpolation algorithm is preferred as the basic model, which has the advantage of quickly capturing spatial gradient changes and is suitable for the characteristics of significant fluctuations in response values within a short distance in mountainous terrain. The input of interpolation is the center coordinates (latitude and longitude or local coordinates) and response index value, and the output is a continuous distribution map of response index covering the entire PV field area. This heat map can clearly show the distribution of high and low response intensity areas, providing a basis for spatial sensitive partitioning. Based on the response index heat map, the system performs array spatial sensitive partitioning operation, and divides the array into high response area, medium response area and low response area by setting response threshold. In order to adapt to complex mountainous terrain, the digital elevation model (DEM) is introduced to correct the terrain gradient based on the preliminary threshold partitioning result. For example, even if the response value is high, the steep slope area may be re-labeled as a low priority area due to high operation cost or low structural stability. The final partitioning result is used to drive the resource allocation logic: the high response area is set as the priority setting target, which means it will be scheduled to perform setting action in the next period; the medium response area is used as the candidate area for dynamic rotation; the low response area is marked as the observation and maintenance area, and the system only performs state monitoring without actively performing setting, saving operation and maintenance resources.

[0047] After completing the regional priority division, in order to achieve precise setting, the system needs to establish a response simulator model for each setting action, that is, to predict the possible power improvement of the target component or area under different types and amplitudes of setting operation. The data sources include past setting behavior and its corresponding short-term power generation effect change data, forming a historical operation-response pair. In the data preprocessing stage, abnormal samples caused by extreme weather or hardware failure need to be removed, and operation parameters (such as angle change amount, cleaning frequency) need to be normalized to improve the generalization ability of model training. For each setting type (such as inclination adjustment), a nonlinear regression model such as support vector regression (SVR) or random forest regressor can be trained separately; for the simulation of strategy cooperation of combined setting actions, a feedforward neural network is introduced, whose input is the setting action vector (combination of multi-dimensional setting parameters), and the output is the predicted combination benefit score. The network structure is usually a three-layer fully connected neural network, the number of hidden layer nodes is set according to the input dimension and sample size, generally using ReLU activation function, the loss function chooses mean square error, and the optimizer uses Adam to improve the convergence speed.

[0048] The training process is based on a historical sample set, and cross-validation is used to ensure the model's generalization ability. After training, the simulator model can accept the current partition state and the combination of tuning candidate parameters as input in the actual deployment stage, quickly evaluate each tuning action or combination strategy, and output the benefit score. The system combines all candidate combinations into a tuning action space, which is used as the optimization object. The optimal tuning scheme under the current constraint conditions is found by a strategy filter (greedy search, genetic algorithm, or reinforcement learning can be introduced). The optimal strategy is fed back to the execution module, which is deployed in the priority area for actual deployment, and the actual power generation change data is again included in the historical sample pool to realize model self-evolution.

[0049] To achieve dynamic tuning optimization and adapt to environmental disturbances and non-static characteristics of system changes, the system needs to be continuously tracked and feedback adjusted in time windows. Specifically, after the high-response area is identified, the system uses a sliding time window mechanism (such as 5 minutes, 15 minutes, 1 hour, etc.) to collect power output data from each component in the area in real time, and correlates and tracks meteorological factors in the corresponding time period, including but not limited to irradiance, 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 (collecting voltage, current, temperature, light, etc.) and meteorological station or micro-meteorological node data accessed through the SCADA platform for unified uploading, cleaning, and management.

[0050] Align the time, i.e., ensure that each power data and meteorological data is organized in a uniform time step; secondly, filter outliers, such as irradiance being 0 but power being greater than a threshold or data jumps caused by inverter failure, and remove unreliable data through statistical analysis (such as median deviation); then perform standardization to eliminate training bias caused by different data dimensions. In addition, low-pass filtering or moving average is used to smooth continuous data, reduce noise interference, and enhance the sensitivity of trend identification. Based on the tracked data, the system builds a time series drift analysis model to determine whether the current power loss path has shifted. For example, the original power decline trend is highly correlated with shading changes, but the shading remains stable and the temperature rises in the current sliding window, causing the output to continue to decline, so it is determined that the dominant loss path has shifted. This judgment relies on correlation analysis, sliding window mutual information calculation, or the use of a micro-LSTM prediction model to evaluate the partial derivative contribution of different factors to power changes. The model structure can be a multi-input single-output neural network, where the input is the change sequence of each meteorological factor in the time window, and the output is the power change prediction residual, from which the control dominance of different factors on the current loss trend is inferred.

[0051] Once the loss path deviation is identified, the system loads the alternative setting strategy matching the new dominant factor, which has formed a mapping relationship in the historical setting data training, for example, when the high temperature is dominant, it tends to introduce an increased ventilation distance, adjust the support angle to reduce the heat absorption area, etc., while the dust is dominant, the cleaning priority strategy is loaded. The setting system will execute the fine-tuning operation from the strategy library according to the current component area, controllable device type and resource scheduling capability. The characteristics of fine-tuning are small amplitude and low interference, such as ± 2° tilt angle adjustment, local MPPT parameter adjustment, local short-term cleaning of specified components, etc. After the fine-tuning operation is implemented, the system immediately enters the short-term evaluation stage.

[0052] The evaluation stage uses the power change data before and after fine-tuning in the sliding window, the change of power output standard deviation (judging consistency), and the change of inverter efficiency (judging conversion ability) to construct a set of setting benefit evaluation indexes. The core evaluation index is the comprehensive short-term benefit function, such as power improvement amplitude, output consistency improvement rate, and operation fluctuation reduction degree, which is usually constructed as a composite weighted function, and the weight is given by the historical strategy effect regression model. If the evaluation function value exceeds the historical benchmark (or model threshold), it is judged that this fine-tuning is an effective strategy, which is included in the effective strategy set under the current environmental state, and its setting weight is improved, and it is recorded into the strategy evolution library for the training data of the subsequent intelligent recommendation system; if the evaluation is not up to standard, the strategy is marked as low priority or temporarily frozen, and feedback to the optimization module to correct its use scene description or parameter boundary.

[0053] Subsequently, the setting system updates the current strategy recommendation ranking according to the strategy benefit results, and preferentially calls the newly verified effective strategy in the next round of setting scheduling, forming a strategy closed loop of exploration-verification-evolution. In the subsequent strategy recommendation model, the system can use a reinforcement learning framework (such as DQN) or a policy gradient method to adjust the behavior output probability according to the historical score of the strategy, and further realize the self-learning and long-term evolution of the setting action, so that the whole system has the intelligent setting ability to adapt to the complex and variable mountain environment. This method breaks through the limitations of traditional periodic fixed strategies and artificial subjective judgment, and truly realizes a real-time data driven, feedback closed loop photovoltaic setting optimization mechanism.

[0054] Embodiment 1:

[0055] A new energy intelligent operation and maintenance engineer Li is responsible for setting optimization of a photovoltaic array located in a typical mountain station in Yunnan, China. The installed capacity of the station is 12MWp, distributed on multiple mountain steps with an altitude difference of 65 meters, and the components are unevenly distributed in height. There are problems such as tree shading, dust deposition and inverter mismatch in local areas, and recently it is found that the total power generation is about 18% lower than the design value. Li decides to use a new generation of array space sensitive partitioning method to intelligently set the system to restore the power generation efficiency.

[0056] First is S1: Obtain operating parameter data. Li calls the SCADA system operating data of the power plant for nearly 30 days, extracts the data of each component in combination with the high-precision DEM modeling result of the unmanned aerial vehicle, including the position coordinates (x, y, z) of the component, the inclination angle (such as 32°-38°), the real-time power output (5-minute granularity), the incident light irradiation (fluctuating between 720-960 W / m 2 2), the surface temperature (45°C-72°C), and introduces the shading degree data (based on three-dimensional terrain modeling and sunshine simulation calculation, the shading value of each component is between 0-0.5) through the shading index model. The data totals more than 110,000 records, which, after time synchronization, missing value interpolation and standardization preprocessing, are used as the data basis for setting the response model.

[0057] Then enter S2: Build response index. Li extracts the past setting behavior data (such as angle adjustment, cleaning frequency, MPPT parameter modification, etc.) and the power improvement amplitude of the component before and after each setting behavior, calculates the average improvement rate under the unit setting influence. For example, a component numbered C2317 in the southeast area brings 6%, 5%, and 8% power improvement in 3 angle adjustments, and its response index is set to After performing this process on all components, a set of discrete response index data is obtained, with the value range being about 1.5-10.8 and the average value being 4.7. These indexes will be used as the original point data for subsequent spatial interpolation.

[0058] Then perform S3: Spatial interpolation modeling. Li selects the inverse distance weighted interpolation (IDW), takes the component coordinates and response index value as input variables, and generates a continuous response index distribution map (heat map) covering the entire power plant site area. To enhance the smoothness of the edge transition, the power parameter p=2.5 in IDW is used, indicating that the closer points contribute more to the interpolation result. During the interpolation process, he finds that the southeast step area and the northwest slope area show obvious high response index aggregation area (greater than 7.0), while the response index near the edge of the shading forest and the periphery of the central inverter area is generally low (less than 3.5). After the interpolation is completed, a high-resolution response heat map is obtained, and the darker the area on the map, the higher the setting potential. The heat map data supports the subsequent strategy deployment.

[0059] Finally, perform S4: Region division and setting execution. Li divides the area according to the response index threshold: R i ≥7.0 is defined as a high response area, 4.5 i <7.0 is defined as a medium response area, and R i≤4.5 is defined as the low response area. After overlaying the partition boundaries on the heat map, it is found that the high response area accounts for 28% of the array area, mainly concentrated in the southeast slope angle inconsistency area and the back slope bias installation area. Subsequently, Li Gong deploys the angle adjustment robot to implement ±3° support adjustment in this area, and arranges the cleaning robot to clean the modules with heavy surface contamination locally, while changing the group string connection mode for several components groups with high response but large current fluctuation from 6 strings in parallel to 4 strings in parallel to match the inverter input voltage curve. The medium response area is included in the next stage of evaluation observation list; the low response area (such as the northwest corner edge, tree shadow interference area) is only set as a monitoring object, without any active adjustment, reducing maintenance costs.

[0060] Within 72 hours after implementation, the system compares and analyzes the power output of the high response area, and finds that the average component power in the adjustment area is increased by 9.4%, which is higher than the 0.8% increase caused by the natural fluctuation of the medium response area; at the same time, the output stability is significantly improved, and the power standard deviation between components is reduced by 22%, and the system current matching degree is improved by 13%. According to the feedback mechanism of the adjustment system, this round of adjustment strategy is marked as an efficient strategy by the system, and the strategy score is increased from 0.61 to 0.87, and is recorded into the strategy recommendation system for future scene matching.

[0061] To further improve the accuracy of the strategy, Li Gong decides to use the response index introduced in this embodiment to build a causal path analysis method, combined with the spatial position and shading mode of the components, to deeply identify the key influencing factors of the actual dominant power loss, and update 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 components and the shading structure. Li Gong retrieves the high-precision digital elevation model (DEM) generated by the unmanned aerial vehicle aerial survey, combined with the GIS geographic information system and the photovoltaic layout map, to accurately calibrate the relative elevation, slope (such as 15°, 22°, etc.), azimuth (such as 10° south of west) and geographic position of each component in the array. He pays special attention to the southeast step area and the two sides of the central channel, where the terrain changes dramatically and the shading mode is complex. By importing the local past 90-day solar trajectory and mountain shadow projection data through the solar simulation software (such as SolarAnywhere), Li Gong establishes the probability function of each component being shaded during the daily sunlight time, and outputs the shading coverage time ratio, such as the shading time of component C1045 accounts for 31% of the total sunlight time, while component C3072 is only 4%.

[0063] To establish the causal relationship between these spatial factors and the response index, Li used a causal inference model for data-driven analysis. He chose to use a structured Bayesian network as the causal path learning model, with input variables including: component height difference Δz, slope θ, orientation α, shading coverage ratio s, historical cleaning frequency f, angle adjustment frequency a, temperature mean T avg , and the response variable being the average power increase ΔP after performing the setting. By using the K2 greedy search algorithm to construct the network structure, Li found that the shading ratio s and the slope θ formed the main path affecting the power increase, with a path confidence of 0.86. Although there was a correlation between the angle adjustment frequency a and ΔP, it was only significant when the slope angle θ was greater than 20°, indicating that angle adjustment was more effective for high slope components and less effective for low slope regions. In addition, the path strength between temperature T avg and ΔP was 0.44, and the path strength between cleaning frequency f and ΔP was 0.38, indicating that temperature had a secondary impact and dust had a third impact.

[0064] Based on this causal structure map, Li adjusted the way the response index was constructed. Originally, the response index was calculated only by the average power increase before and after setting (e.g., component C1101 response index was 6.7), now the key variables in the causal path were added to the weight correction. 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 shading ratio, θ i is the slope angle, and the weights are determined by the edge strength normalization in the Bayesian network, ω s = 0.5, ω θ = 0.3. Substituting the data for component C1101: ΔP = 6.7%, s i = 0.24, θ i = 21°, the corrected response index is:

[0067]

[0068] For another component C3072, ΔP is 6.2%, but s i is only 0.04, and θ i is 9°, its corrected response index is:

[0069]

[0070] This means that although the two components have similar elevation before and after the setting, the response index after the causal correction shows a significant difference. The system will prefer the area where C1101 is located for subsequent setting, and C3072 will be included in the sub-optimal level. This causal weight correction mechanism effectively solves the setting bias problem of strong surface response but unexplained root cause.

[0071] After applying the updated response index model, Li Gong regenerated the array heat map and adjusted the spatial sensitive partition boundary. The result shows that some of the original medium response areas are reclassified as high response areas due to severe terrain obstruction and steep inclination. The strategy scheduling system is also updated. The subsequent setting effect is significantly improved. After adjusting the support angle and removing the obstruction, the average power of the reclassified high response area is increased by 12.3% within 24 hours, which is much higher than the original model estimate of 7.5%. The system finally evaluates that the causal weighting modeling mechanism improves the efficiency of the setting strategy by about 65%.

[0072] The partition method based on response index spatial interpolation and modified response potential function model proposed in this embodiment accurately identifies the area with the highest setting benefit potential by constructing a continuous spatial distribution function Ψ(x, y) that reflects the potential of component setting. Especially in complex mountainous environments, this method has a significant advantage.

[0073] First, in this round of optimization, Li Gong collected real-time running data for all array components, including the output power of a unit component (unit W), current (unit A), irradiance intensity (unit W / m 2 ), component temperature (unit ℃), and installation location coordinates (latitude and longitude and altitude z). The power station has deployed 13,200 components, and each component records data at a 10-minute interval, forming up to 2 million historical running records. To extract the setting potential, Li Gong calculates the response index R i of each component, which is the average unit setting gain of the component after historical setting behaviors (such as inclination adjustment, local cleaning, and electrical reorganization), defined as follows:

[0074]

[0075] Where ΔP i,j is the power increment (unit W) brought by the jth setting, A i,j is the corresponding setting amplitude (such as angle change or cleaning area), and N i is the number of historical settings. Li Gong selects the component C 0537 response data as an example: three settings bring 15W, 12W, and 14W improvements, with a setting amplitude of 2.5°, 1.8°, and 2.2°, respectively.

[0076]

[0077] Subsequently he will all components of R i The value of its (x, y) coordinates, select the inverse distance weighted interpolation (IDW) algorithm to generate the preliminary response index heat map R (x, y). The method sets the power index p = 2.2, the response value of each point (x, y) weighted average, the closer the higher the weight of the components. In the interpolation process, Li found that the components located in the south of the response index is generally higher (6.5 ~ 8.9), and the north of the serious area of shelter is less than 3.0. To enhance the image smoothness, he compared between spline fitting and IDW, and finally because the IDW processing slope boundary is more sharp, select it as the main algorithm.

[0078] After obtaining the continuous map of R (x, y), Li introduced the revised response potential function model of the present embodiment:

[0079]

[0080] Where θ (x, y) is the angle between the component orientation and the solar incident angle at noon on the day (in radians), which is simulated by matching the geographical orientation and solar azimuth data. If the component orientation is 10° west of south, and the solar azimuth is due south, the angle is θ = 10° = 0.1745 radians. The terrain gradient The contour slope change is calculated using DEM data to calculate the local slope, with a unit of percentage (such as 26% for a more dramatic area).

[0081] Li sets the correction coefficient for the model: α = 0.4, used to amplify the penalty effect of the angle deviation; β = 0.6, used to regulate the penalty weight of the terrain dramatic fluctuation on the strategy failure. These two values are determined according to the historical strategy effect regression fitting, and the recommended use range is α ∈ [0.2, 0.6], β ∈ [0.4, 0.8]. Taking the point where component C 0537 is located as an example, its R (x, y) = 6.34, θ (x, y) = 0.1745, (i.e. 23% slope), then:

[0082]

[0083] Li generates the revised response potential heat map of the entire field Ψ (x, y) value, and divides the area according to the threshold Ψ th = 5.5, higher than this value is the high response area, used for the key deployment of this round of setting, accounting for about 31% of the area; 4.0 ~ 5.5 is the medium response area, enters the strategy rotation pool; lower than 4.0 is the low response area, included in the passive observation and remote monitoring group, no artificial or equipment intervention is arranged.

[0084] Finally, Li Gong performs fine tuning actions in the high response area: component tilt angle adjustment ±2°, local dirty area starts cleaning robot to perform fine trajectory cleaning with a bandwidth of 22 cm, and optimizes the wiring method for the area with significant line loss, and adjusts the slope parameter of the inverter maximum power point tracking (MPPT) curve. 12 hours after the tuning, the real-time monitoring data shows that the average power of the high response area is increased by 11.2%, which is higher than the 3.1% of the medium response area, the standard deviation of the output power is decreased by 19%, and the overall power generation efficiency is increased by 8.7% compared with the previous year. The tuning feedback data is automatically written into the response index model, and the system enters a new round of adaptive correction.

[0085] Due to the large area of the high response area and the limited tuning resources (such as cleaning robots, adjustable angle support controllers, operation and maintenance time windows, etc.), Li Gong uses the power gain prediction model proposed in this embodiment to deduce the tuning sequence, simulates the power generation revenue effect of different tuning combination strategies in each response sub-area, and finds out the tuning priority sequence with the maximum gain and the best synergy, realizing intelligent regulation and control.

[0086] First, Li Gong divides the high response area into 6 sub-areas (Z1-Z6), each with an area of about 150-300 square meters, composed 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 the Z3 area, it is 6.41, in the Z4 area, it is 6.12, in the Z5 area, it is 5.95, and in the Z6 area, it is 6.05. The tuning actions that can be performed in each area include: A1-tilt angle adjustment (±2-3°), A2-surface cleaning, A3-wiring structure optimization, A4-inverter parameter adjustment. Since these tuning actions may be performed in some areas, Li Gong needs to determine whether to do A1 first or A2, or whether to perform A3 only on part of the area to achieve the optimal gain. For this purpose, Li Gong uses the power gain prediction model proposed in this embodiment. The model structure is a multi-input single-output feedforward neural network (FNN), and the input vector includes: the corrected response potential value Ψ of the current area, the average irradiance value of the components, the power difference before and after the last tuning, the planned tuning action type and amplitude (such as tilt angle adjustment 2.5°, cleaning level = 2, etc.), and the output is the expected unit power increase (unit W / m 2 ) brought by this tuning action combination in the area. The network has 3 layers, the first layer is the input layer (7 neurons corresponding to the above 7 input features), the two hidden layers are 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 square error), and the optimizer is Adam. The training set uses a total of 80,000 tuning actions and effect data from the past two years in the power plant.

[0087] Taking the Z3 area as an example, Li Gong tests the following four tuning action combination strategies:

[0088] Scheme 1: Perform A1 only (tilt adjustment 2°)

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

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

[0091] Scheme 4: Perform A1+A3+A4 (full set of actions)

[0092] Substitute the input features of Z3 region into the model: Ψ = 6.41, the average irradiance value is 890 W / m 2 , the power improvement before and after the past setting is 6%, the gain prediction value of each combination is simulated as follows:

[0093] Scheme 1: expected to improve 12.3 W / m 2

[0094] Scheme 2: expected to improve 17.6 W / m 2

[0095] Scheme 3: expected to improve 14.2 W / m 2

[0096] Scheme 4: expected to improve 18.1 W / m 2

[0097] Although the benefit of scheme 4 is slightly higher, it involves all the setting actions, high device resource occupation, and long operation time. The system built-in resource weight model calculates the operation load index (represented by device working hours, control system load, and scheduling cost) required for unit power improvement as follows:

[0098] Scheme 1: 1.8

[0099] Scheme 2: 2.9

[0100] Scheme 3: 3.1

[0101] Scheme 4: 4.5

[0102] Divide the gain value by the operation load index to obtain the cost-effective index:

[0103] Scheme 1: 6.83

[0104] Scheme 2: 6.07

[0105] Scheme 3: 4.58

[0106] Scheme 4: 4.02

[0107] Finally, the system decides to execute scheme 1 first, and performs angle fine-tuning for Z3 area. Z1 area is selected to execute scheme 2, which is a combination of cleaning and angle adjustment, because its response potential is lower than that of Z3 but slightly higher than that of Z5. Z2 and Z5 areas are selected to execute A4, which is to adjust MPPT parameters only, because their slopes are high and cable laying is complex. Z6 area is selected to execute A3 (local series-parallel replacement) because its wiring topology is complex and its response performance is good in the early stage. After the strategy selection is completed, Li Gong executes the operations in order, and the system records the power improvement data after 24 hours. The power of Z3 area is improved by 12.9 W / m 2 , which is slightly higher than the model prediction value of 12.3, verifying the accuracy of the model prediction. At the same time, the system automatically adds the actual tuning data of this time to the model training data set for subsequent model retraining, improving the long-term learning ability.

[0108] Example 2

[0109] Based on example 1, in a certain mountain photovoltaic power station in Honghe, Yunnan, with an altitude of 1650 meters, the response potential function modeling, response area division and strategy priority arrangement have been completed. In continuous operation, the system detects that the component power of Z5 area (northwest slope) abnormally fluctuates, which is about 11.6% lower than the average of one week ago. Li Gong decides to start the tuning and evaluation of short-term income of the tuning scheme proposed in this embodiment, and executes it according to the technical path of S1-S4.

[0110] S1: Real-time acquisition of operating parameters. Li Gong uses the photovoltaic monitoring system SCADA and string acquisition terminal to collect real-time data of 142 components in Z5 area. The collected indicators include: current power output of the component (unit W), current (unit A), voltage (unit V), component surface temperature (℃), solar radiation (W / m 2 ), shading index (based on shading image recognition score, between 0 and 1), wind speed, humidity and environmental temperature. Taking C507 component as an example, its current power is 174 W (about 14% lower than the design value), current is 4.5 A, component temperature is 68℃, shading index is 0.08, light intensity is 910 W / m 2 , and wind speed is 2.1 m / s.

[0111] S2: Build multi-period performance trends and identify dominant factor shifts. Li Gong calls the component performance data and operation logs of the power plant in the past 72 hours, compares it with the same period last month, builds a three-section time window (short cycle 3 hours, medium cycle 24 hours, long cycle 72 hours) trend model, and uses sliding window correlation analysis method. The system determines that the light of most components in Z5 area has not fluctuated significantly (the correlation remains above 0.92), the shading state is also relatively stable, but the negative correlation between temperature change and power drop has risen from -0.45 to -0.78, and the system identifies that the power loss dominant factor shifts from shading to temperature impact. Li Gong calls the preset rules accordingly: as the summer high temperature approaches 75℃, some components are close to the performance degradation threshold (about 70℃), and temperature suppression type tuning strategy needs to be loaded.

[0112] S3: Load alternative tuning strategy and implement fine tuning. The system automatically calls the tuning template for high-temperature dominant loss path, and suggests implementing a light-weight ventilation-enhanced support angle optimization strategy in the affected area, i.e. a slight angle adjustment of 1.5° to increase back air flow, and conducting a spot water curtain spraying experiment to test its cooling effect. Li Gong selects components C507, C510, C514, and C520 in Z5 area as samples to implement fine tuning, adjusts the angle from 32° to 33.5°, and starts the spraying system for 1 minute / interval 10 minutes spraying cycle, lasting for 30 minutes.

[0113] S4: Short-cycle evaluation of fine-tuning benefits. Within 1 hour after adjustment, the system continuously collects component parameters such as C507 at 5-minute granularity, and evaluates indicators including:

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

[0115] Current consistency (σ I ): The original standard deviation is 0.83A, which decreases to 0.61A, and the consistency is improved by 26.5%;

[0116] System load balancing degree (LoadIndex): The power fluctuation amplitude of the input side of inverter Z5 decreases from ±12.6% to ±6.2%.

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

[0118]

[0119] Where the weights are set as w1=0.5, w2=0.3, w3=0.2, and the substitution gives:

[0120] A = 0.5*0.0977 + 0.3*0.265 + 0.2*0.5079 = 0.04885 + 0.0795 + 0.1016 = 0.22995

[0121] The result is higher than the fine-tuning strategy benefit threshold of 0.15 set locally, and the system judges that the fine-tuning strategy is effective, and the ventilation type angle up and the local spray strategy are included in the current main recommended strategy set of Z5 area, and the priority is promoted to class A. At the same time, the system stores the fine-tuning operation and evaluation data into the strategy training log for subsequent regression model optimization, and sets the component group as the strategy evolution point, allowing the system to continuously optimize its setting mode through self-learning in the future.

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

[0123] The embodiment establishes a setting strategy library suitable for various dominant loss factors, and deploys an intelligent setting fine-tuning system that can be called and executed in real time. The core of the system is the training process of the strategy-scene mapping model and the automatic execution ability of the operable action module.

[0124] First of all, Li Gong organized the operation and maintenance team to analyze the common power loss types in the operation of the power station in the past two years, and divided them into five categories of dominant factors: ①shading dominant type, ②high temperature dominant type, ③pollution and dust accumulation dominant type, ④electrical mismatch dominant type, and ⑤seasonal light deviation dominant type. For each main factor, the system constructs a strategy template library, and each template record contains three types of information: (1) applicable condition feature set; (2) recommended action combination set; (3) execution parameter suggestion. For example, the high temperature dominant type template contains environmental temperature > 32°C and component surface temperature > 65°C, daily irradiance > 850W / m 2, shading index <0.15, wind speed <3 m / s; recommended actions include tilt angle up 2-3°, local ventilation optimization, MPPT response curve fine-tuning. After the template is formed, Li Gong uses the existing 80,000 historical setting and operation data of the power station to match and label, as subsequent classification model training samples. All historical samples are summarized by hour granularity to construct sample feature vector X, including: environmental temperature, component temperature, shading index, slope, wind speed, irradiance, historical power deviation rate, sunshine duration deviation, etc. 8-dimensional indicators. Label Y is the actual execution strategy type (such as Template_2 represents high-temperature setting). For this multi-class classification problem, Li Gong uses a lightweight gradient boosting tree (LightGBM) to build a classification model, which outputs the predicted strategy type number after inputting the feature vector. The model structure has 200 trees and a maximum depth of 8, and the average accuracy of 10-fold cross-validation reaches 92.3%, which is better than the traditional logistic regression model. After the model is trained, whenever the system identifies a power loss main cause shift (such as from shading to temperature), the model will call the matching strategy template according to the current component environmental conditions and generate an operation list in real time.

[0125] In practical applications, Li Gong selects high-temperature setting fine-tuning for Z3 area in the system, and the recommended strategy output by the model is Template_2, and the specific fine-tuning actions are as follows:

[0126] 1. Component tilt angle fine-tuning: within the adjustable range of the support ±5°, the angle is adjusted to 33.5° to improve the back convection intensity. The system reads the recommended angle from the strategy template library, which is 2° higher than the current base, and sends the angle adjustment command to the corresponding group through the PLC control support actuator, with an adjustment accuracy of ±0.2°.

[0127] 2. Local switching of module string connection mode: identify that the original layout of this area is 6 strings in parallel connection mode, which causes some current weak components to drag down the output of the whole string. The system recommends changing to 3 strings in parallel connection mode on specific branches to improve overall voltage matching efficiency. The electrical switching scheme is completed through the intelligent circuit breaking system in the programmable photovoltaic combiner box, and the operation is remotely executed after authorization by the background system.

[0128] 3. Point cleaning operation: call the pollution tolerance threshold recorded in the strategy library (surface reflectance decreases by more than 15%), and combine the latest unmanned aerial vehicle infrared imaging data to find hot spot aggregation areas C316 and C319, arrange the point low-pressure water jet robot to clean, with a spray intensity of 1.2 L / min and a duration of 3 minutes, which belongs to the light cleaning level.

[0129] After the fine-tuning operation is completed, the system records the execution time, operation parameters, execution area, component response, and short-term power gain within 1 hour of the tuning action, and constructs a strategy effect log. If a certain strategy is continuously evaluated as efficient in the same scenario, its strategy weight will be increased, and it will appear in the recommended strategy sequence in the future. On the contrary, if an action performs poorly in the same scenario, it will be automatically adjusted to a lower priority by the system and added to the strategy review queue for subsequent model iteration for manual review and feature adjustment.

[0130] The multi-dimensional operation data is used to make an immediate value judgment on the tuning behavior, and to determine whether to increase its weight in the future recommended strategy. This evaluation is carried out around 116 components in the Z3 area. Li Gong uses 5-minute granularity sampling to continuously collect key operation parameters within a 60-minute window after the tuning is completed, and calculates the short-term benefit value Λ according to the set benefit function. t .

[0131] First, Li Gong calculates the average power increase amplitude ΔP t . Before tuning, the average unit component power in the Z3 area is 172.4W, and the running data within 15-60 minutes after tuning shows a stable increase to 186.3W, so:

[0132] ΔP t = 186.3-172.4 = 13.9W

[0133] Second is the output consistency improvement rate Σ t . Before tuning, the power standard deviation of 116 components in the Z3 area is 21.8W; after tuning, it decreases to 16.5W, and the consistency improvement rate is:

[0134]

[0135] Third is the system current fluctuation rate Ω t . Li Gong obtains the maximum current change amplitude of Z3 branch through real-time current sampling of the inverter. Before tuning, the current fluctuation standard deviation is 2.65A, and after tuning, it decreases to 1.93A, so:

[0136]

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

[0138]

[0139] Li Gong sets the experience adjustment factor in the system to γ = 0.6, η = 0.7, which is determined through a large number of scene simulation and strategy verification, and the recommended range is γ ∈ [0.4, 0.8], η ∈ [0.5, 1.0]. Substitute the values into the formula:

[0140]

[0141] The system sets a reference threshold for the effectiveness of the setting strategy as Λ th = 8.0, and determines it as a high-efficiency strategy when the value exceeds the threshold. Since Λ t = 10.54 > 8.0, the system evaluates the composite setting strategy of this time's inclination adjustment + cleaning + wiring structure optimization as effective, and marks it as a priority strategy in the strategy recommendation system. The system then performs two feedback actions: one is to write the strategy into the strategy recommendation priority table of the current station, so that when similar environmental conditions occur in the future (such as high temperature in summer and large temperature difference between components), the strategy will be prioritized for scheduling; the second is to record the strategy operation steps, parameters, timeliness and execution effect in the strategy learning library, which is used for training of the subsequent deep reinforcement learning model and self-evolution of the strategy library.

[0142] In addition, to verify the generalization of the setting action, Li Gong replicates and implements the same strategy in Z4 area and collects data for comparison. The power improvement in Z4 area is slightly lower, with an average of 10.2 W, a consistency improvement rate of 0.19, and a fluctuation rate of 0.81. After substituting the same parameters, we calculate:

[0143]

[0144] which is lower than the set threshold, and the system automatically marks the strategy as context-dependent, i.e., not applicable in all areas and reduces its generalization recommendation weight. This mechanism ensures that the strategy is only applied in areas with structural benefit potential, avoiding one-size-fits-all setting that leads to resource waste.

[0145] In summary, this example fully demonstrates the practical feasibility of the short-term setting benefit evaluation mechanism proposed in this embodiment, especially through real-time collection of power, current and consistency indicators, and construction of a composite benefit function, which can accurately identify the value of the strategy. At the same time, with the application of dynamic adjustment coefficient, the method has scene self-adaptation ability. More importantly, the structure of the benefit value Λ t not only has interpretability, but also can be used in subsequent intelligent systems such as reinforcement learning and strategy scoring, forming a three-dimensional closed loop of strategy-data-effect, and is one of the key basic modules of future intelligent photovoltaic setting systems.

Claims

1. A method for setting the power loss of a mountain photovoltaic array, characterized in that: a causal graph of the coupling of multiple sources including terrain, angle, sunshine, temperature, dust, and wiring mode to component output is established; a Bayesian network-based identification method identifies the dominant loss path; a structural weight matrix of output power loss is identified, including the influence factor group of inclination > temperature > wiring topology on the setting; based on the loss path weight, the power loss response index of each component is calculated; a response index heat map is formed using a spatial weight interpolation algorithm to perform array spatial sensitive partitioning; the high response area is used as the priority setting target, and the low response area enters the observation and maintenance mechanism; based on historical data, a simulator is trained for each setting action to build a setting action space, and a nonlinear regression or neural network is used to evaluate the synergistic benefits of combined strategies; in time windows, the power output changes in the high response area and their drift trends with meteorological factors are continuously tracked; if the system identifies a loss path shift, the corresponding alternative setting strategy is loaded; fine tuning is implemented, and its short-term benefits are evaluated; if the effect is good, a new round of strategy iteration is entered; the method for performing array spatial sensitive partitioning comprises: S1, obtaining the operating parameter data of each photovoltaic component in the mountain photovoltaic array, including power output, shading degree, inclination, terrain coordinates, solar radiation, and temperature information; S2, calculating the power variation amplitude of each component under historical setting behavior, and constructing a component response index based on it to reflect the actual power response capability to setting measures; S3, using a spatial interpolation algorithm to model the discrete response index spatially continuously to form an array response index heat map; S4, dividing the entire photovoltaic array into high response areas, medium response areas, and low response areas; identifying the high response area as the priority setting target and executing angle adjustment, component cleaning, and electrical connection optimization setting operations; executing observation and maintenance strategies for the low response area without implementing resource setting operations.

2. The method of setting the power loss of a mountain photovoltaic array according to claim 1, wherein During the response index construction process, the spatial location and shading mode of the component are combined to establish a causal influence path to identify the impact factors of power loss.

3. The method of setting the power loss of a mountain photovoltaic array according to claim 2, characterized in that The spatial interpolation algorithm includes, but is not limited to, 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 setting of the response index threshold, and the priority division supports correction based on terrain contour distribution to adapt to the complex topographic features of the mountain.

4. The method of setting the power loss of a mountain photovoltaic array according to claim 3, wherein The priority execution order of the setting operation is determined by a power gain prediction model to evaluate the synergistic gain effect of different setting actions on each response area.

5. The method of setting the power loss of a mountain photovoltaic array as claimed in claim 1, wherein The method for implementing fine tuning and evaluating its short-term benefits comprises: S1, real-time acquisition of the operating parameters of each component in the mountain photovoltaic array, including power output, current, voltage, component temperature, light intensity, shading state, and external environmental factors; S2, based on the operating parameters, a multi-cycle performance trend relationship is constructed to monitor the dominant influence factors of power loss; when a dominant factor migration is identified in the current power loss path compared to the historical operating state, it is determined that the loss path has shifted. S3, for the identified dominant factor, load the corresponding alternative setting strategy template, including the inclination adjustment, component cleaning, wiring structure adjustment or regional power scheduling; small range setting fine-tuning operation is implemented on the identified area; S4, the power gain, current consistency and system load balance degree index of the setting area are evaluated in a short period after fine-tuning, and a short-term setting benefit vector is constructed; if the evaluation result shows that the setting strategy is effective, the strategy is included in the subsequent setting mainstream process; if it is not effective, it is recorded in the strategy correction library for subsequent relationship optimization.

6. The method of setting the power loss of a mountain photovoltaic array according to claim 5, wherein The alternative setting strategy template pre-establishes a matching strategy library for different loss main factors, and calls according to the identification result before setting; the fine-tuning operation includes component inclination adjustment within ±5°, local switching of string wiring mode or short-time fixed-point cleaning operation of target components.

7. The method of setting the power loss of a mountain photovoltaic array according to claim 6, characterized in that The short-term setting benefit evaluation is based on the comprehensive judgment of the power improvement amplitude, the component output consistency improvement rate and the system operation stability after setting; when the setting strategy is verified to be effective and repeatedly appears, the output weight is improved, and it is preferentially included in the recommended strategy set of the subsequent setting process.

Images