Self-adaptive cleaning decision-making method and system for photovoltaic power station in resource unconstrained scene

Abstract

The invention discloses a self-adaptive cleaning decision-making method and system for a photovoltaic power station in a resource unconstrained scene. Aiming at a one-machine one-array resource unconstrained scene, the method comprises the following steps: performing distributed sensing by taking a single photovoltaic array and an exclusive cleaning robot thereof as independent units; constructing an independent income-cost dynamic starting threshold model for each unit; when the pollution rate exceeds a threshold value, a net cleaning utility value is calculated based on refined weather forecast, if the net cleaning utility value is smaller than a mud spot risk threshold value, cleaning is triggered immediately, and if the net cleaning utility value is larger than the cleaning cost, a rainwater replacement strategy is executed; when cleaning is needed, in combination with future time-of-use electricity price curve optimization, the moment with the maximum expected net income is selected to start operation; and after cleaning is completed, performing closed-loop feedback and iterative updating on the model reference coefficient based on actual energy efficiency improvement data. According to the invention, by introducing the rainfall effectiveness dynamic game, the electricity price window active optimization and the online self-learning correction mechanism, the intelligent and adaptive optimization of the cleaning decision is realized.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant operation and maintenance technology, and in particular to an adaptive cleaning decision-making method and system for photovoltaic power plants based on dynamic thresholds, multi-factor game theory and closed-loop learning in a scenario with unconstrained resources (i.e., "one machine, one array") Background Technology

[0002] Driven by the strategic goal of "carbon peaking and carbon neutrality," the installed capacity of photovoltaic power generation continues to grow. However, the deposition of pollutants such as dust and bird droppings on the surface of photovoltaic modules can lead to a serious "shading effect." Studies have shown that such pollution can reduce the power generation efficiency of power plants by 17% to 40%, seriously affecting economic benefits.

[0003] Currently, research on intelligent cleaning decision-making in photovoltaic power plants is mostly based on "resource-constrained" scenarios, where limited cleaning equipment, such as mobile robots or cleaning teams, needs to be scheduled among a large number of photovoltaic arrays. Such problems are typically modeled as the Traveling Salesman Problem or the Vehicle Routing Problem, and solved using heuristic algorithms such as particle swarm optimization and genetic algorithms. While these methods are theoretically feasible, they face challenges in practical engineering, including high computational complexity, sensitivity to model parameters, and weak real-time decision-making capabilities, limiting their large-scale application.

[0004] In recent years, with the popularization of the "one machine, one array" model, resource constraints have been lifted, and the core of decision-making has shifted to finding the optimal cleaning time for each independent array to achieve global benefits. However, existing solutions are mostly based on simple rules, such as fixed cycles, single pollution rate thresholds, or static multi-factor combinations, which have obvious defects: (1) They cannot accurately quantify the "double-edged sword" effect of natural rainfall, as light rain may lead to a "mud spot effect" that reduces power generation; (2) They fail to dynamically optimize returns in conjunction with time-of-use pricing, and simply avoiding high-price periods may result in missing higher return opportunities; (3) They lack self-learning and correction capabilities, and the fixed model parameters cannot adapt to differences in array characteristics and long-term environmental changes. Therefore, there is an urgent need for an adaptive decision-making scheme that can engage in dynamic game theory, optimize returns, and evolve itself. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an adaptive cleaning decision-making method and system for photovoltaic power plants in resource-unconstrained scenarios. This invention not only achieves distributed independent decision-making but also solves the problems of inaccurate cleaning decisions, poor economic efficiency, and lack of adaptability in existing technologies by introducing an effective rainfall game mechanism, a price window revenue optimization model, and a closed-loop correction mechanism based on actual results.

[0006] In a first aspect, the present invention provides an adaptive cleaning decision method for photovoltaic power plants under resource-unconstrained scenarios, comprising the following steps:

[0007] S1: Distributed sensing and data acquisition: Taking a single photovoltaic array and its dedicated cleaning robot as an independent decision-making unit, real-time pollution rate data, weather forecast data, time-of-use electricity price information, and equipment status data of the cleaning robot are collected in real time.

[0008] S2: Dynamic Start-up Threshold Calculation and Pollution Determination: Calculate the dynamic pollution rate threshold θ for the independent decision-making unit. S(t) And compare the real-time contamination rate with the dynamic contamination rate threshold θ S(t) If the real-time contamination rate does not exceed the dynamic contamination rate threshold, the decision is to remain on standby.

[0009] S3: Effective Rainfall Game Theory Determination: If the real-time pollution rate exceeds the dynamic pollution rate threshold θ S(t) The system obtains a refined weather forecast (including the probability of rainfall and the predicted rainfall amount) for a future preset time window (e.g., 24-48 hours), and calculates the net cleaning utility value of the predicted rainfall. The core of this calculation lies in the quantitative assessment of the risk of mudslide effect caused by rainfall. When the net cleaning utility value is less than the preset mudslide risk threshold, it is determined that waiting for rainfall is uneconomical or risky, the rainfall factor is ignored, and an instruction to immediately execute machine cleaning is generated.

[0010] S4: Electricity Price Window Optimization: After determining that machine cleaning is necessary, instead of immediately or simply delaying it to a low-price period, the system actively optimizes revenue by combining the time-of-use electricity price curves for the next N hours (e.g., the next 24 hours). The system calculates the expected net revenue of immediate cleaning and the expected net revenue of cleaning during potential future low-price periods. The expected net revenue comprehensively considers the generation loss during the delay period, the price differences between different periods, and the cleaning cost. The system selects the moment with the largest expected net revenue as the final operation start point and generates the corresponding cleaning instruction.

[0011] S5: Adaptive Closed-Loop Correction: After the cleaning operation is completed, based on the actual operating data fed back by the dedicated cleaning robot (such as power data before and after cleaning), the actual energy efficiency improvement gain G_real before and after cleaning of the array is calculated, and the deviation between the actual energy efficiency improvement gain G_real and the cleaning gain G_pred estimated when making the cleaning decision is calculated (this deviation reflects the accuracy of the model prediction). According to the deviation, the benchmark coefficient k0 in the dynamic pollution rate threshold is iteratively updated through an online learning algorithm (such as gradient descent method, reinforcement learning algorithm). The updated benchmark coefficient is used for the next decision cycle of the independent decision unit, so that the system can continuously self-optimize and adapt to changes in the environment and component performance.

[0012] Furthermore, the dynamic pollution rate threshold θ S(t) The calculation formula is determined by a combination of cleaning costs, expected power generation revenue, seasonality, and weather factors:

[0013]

[0014] Where k0 is the baseline coefficient updated through online learning, reflecting the system's basic benefit-cost break-even point; OC i P_avg(t) is the cost of a single cleaning operation for the i-th photovoltaic array; P_avg(t) is the predicted average electricity price over a future cleaning effect period; P_irated is the rated power of the photovoltaic array; k_season is the seasonal adjustment coefficient set based on the abundance of solar resources at the location of the photovoltaic array; k_weather is the weather adjustment coefficient set based on the continuity of microenvironment weather forecasts for the photovoltaic array.

[0015] Furthermore, the seasonal adjustment coefficient k_season takes a negative value in summer to encourage early cleaning and a positive value in winter to increase the threshold.

[0016] Furthermore, the weather adjustment coefficient k_weather takes a negative value when forecasting consecutive sunny days to lower the threshold, and takes a positive value when forecasting intermittent dusty or rainy days to raise the threshold.

[0017] Furthermore, in step S3:

[0018] When the forecast rainfall is below the first threshold (e.g., 5 mm), the net cleaning effectiveness is determined to be negative (because mud spots may form).

[0019] When the forecast rainfall exceeds the second threshold (e.g., 20 mm), the net cleaning utility value is determined to be positive, indicating a significant positive cleaning effect.

[0020] When the forecast rainfall is between the first and second thresholds, the probability of mud spot effect is assessed and the net cleaning utility value is calculated based on the historical data model.

[0021] Furthermore, the future N hours mentioned in step S4 are 24 hours, and the formula for calculating the expected net income is: Expected net income = (expected power generation gain after cleaning × electricity price for the corresponding period) - cleaning cost - power generation loss caused by pollution during the delay period.

[0022] Furthermore, the meteorological forecast data is gridded meteorological data containing rainfall forecasts within 72 hours, and the time-of-use electricity price information is obtained in real time through the power grid marketing system interface.

[0023] Secondly, the present invention provides an adaptive cleaning decision system for photovoltaic power plants in resource-unconstrained scenarios, comprising:

[0024] The sensing module is used to collect pollution rate data, weather forecast data, time-of-use electricity price information, and equipment status data of the corresponding dedicated cleaning robot in real time for each independent decision-making unit, on a per-photovoltaic array basis.

[0025] The decision module, which is communicatively connected to the perception module and deployed on an edge computing node close to the photovoltaic array, is used to independently execute steps S2-S5 of an adaptive cleaning decision method for photovoltaic power plants under resource-unconstrained scenarios for each independent decision unit, and generate cleaning instructions.

[0026] The execution module, which is communicatively connected to the decision module, is used to receive the cleaning instructions and dispatch the corresponding dedicated cleaning robot to perform the cleaning operation, and to feed back the operation results to the perception module to form a closed loop.

[0027] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of an adaptive cleaning decision method for a photovoltaic power station under a resource-unconstrained scenario.

[0028] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of an adaptive cleaning decision method for a photovoltaic power station under a resource-unconstrained scenario.

[0029] The beneficial effects of this invention are as follows:

[0030] 1. Intelligent and refined decision-making: By effectively determining rainfall through game theory, the problem of mud spot effect in the utilization of natural rainfall is solved, and the accurate assessment and utilization of rainfall resources is realized.

[0031] 2. Maximizing economic efficiency: By optimizing electricity price windows, the simple time-period avoidance is transformed into a proactive search for maximizing profits, which significantly improves the economic efficiency of cleaning operations.

[0032] 3. Self-evolution capability: An adaptive closed-loop correction step is added, enabling the decision-making model to continuously optimize online based on historical performance. It has self-learning and adaptive capabilities that traditional static models do not have, resulting in better long-term operating efficiency.

[0033] 4. Distributed robustness: The distributed decision-making architecture of "one machine, one array" combined with edge computing avoids the bottleneck of central computing, and the system has strong scalability and high fault tolerance.

[0034] 5. Significant comprehensive benefits: Experiments show that this invention significantly reduces the number of ineffective cleaning operations, improves power generation revenue through precise decision-making, and reduces electricity costs by optimizing electricity prices, thus achieving dual optimization of operation and maintenance costs and power generation revenue, resulting in a significant increase in annual net income. Attached Figure Description

[0035] Figure 1 This is a block diagram of the architecture logic of the adaptive cleaning decision system provided in this embodiment of the invention;

[0036] Figure 2 This is a flowchart of the decision-making method provided in the embodiments of the present invention. Detailed Implementation

[0037] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0038] This application's embodiments are applicable to "resource-unconstrained" scenarios where each photovoltaic array is equipped with a dedicated cleaning robot. For example... Figure 1 As shown, the system consists of a closed loop of a perception module, a decision-making module, and an execution module.

[0039] The perception module is responsible for multi-source data acquisition. It obtains the pollution rate S(t) in real time through the high-precision optical pollution monitors of each array; it obtains gridded weather data containing accurate rainfall forecasts for the next 72 hours through the meteorological API; it obtains time-of-use electricity price tables from the power marketing system; and it collects the power consumption and fault status of each robot through the Internet of Things gateway.

[0040] The decision-making module is the core, deployed on the edge computing nodes of the power plant. It iterates through all arrays at fixed intervals (e.g., every 15 minutes), executing the decision-making process. (See...) Figure 2 For each array i:

[0041] Step 1: Status Check. Check the status of its dedicated robot (such as whether the battery is sufficient and whether there are any malfunctions). If it is unavailable, skip this step.

[0042] Step 2: Dynamic Threshold Calculation and Initial Judgment. Calculate the current dynamic contamination rate threshold θ for this array. S(t) The initial value of k0 is set empirically and will be updated online via step S5. P_avg(t) is calculated based on the time-of-use electricity price forecast for the next 24 hours. k_season: set to -0.02 in summer (abundant sunshine) and +0.01 in winter. k_weather: set to -0.01 if the forecast is for three consecutive sunny days; set to +0.005 if the weather is changeable and there is intermittent dust. Compare the real-time pollution rate S(t) with θ. S(t) If S(t) ≤ θ S(t) If the output is "Stay put", then the output will be "Stay put".

[0043] Step 3: Determining Effective Rainfall in the Game. If S(t) > θ S(t) Then you can query a more detailed weather forecast.

[0044] Example: The forecast indicates that there will be rainfall in 12 hours, with a 90% probability and a forecast rainfall of 8 mm.

[0045] The system calculates that 8mm falls between the first threshold (5mm) and the second threshold (20mm). Based on historical data models, the system assesses that this rainfall has a 30% probability of causing localized mudslides, with a calculated net cleaning utility of +0.5 (unit benefit). The preset mudslide risk threshold is +1.0 (meaning that a significant positive benefit must be generated before it is worth waiting).

[0046] Decision: Since +0.5 < +1.0, the system determines that the utility of this rainfall is insufficient and poses a risk, ignores the rainfall, and proceeds to the next step of optimizing the electricity price. If the forecast rainfall is 25mm and the calculated utility value is +3.0, the system may decide to "wait for rainfall" and set a scheduled task for reassessment after the rain.

[0047] Step 4: Optimal electricity price window. After confirming that machine cleaning is required, the system obtains the electricity price curve for the next 24 hours (e.g., currently at peak time of 1.2 yuan / kWh, entering off-peak time of 0.3 yuan / kWh in 4 hours).

[0048] Calculate the expected net benefit A from immediate cleaning: (Expected power generation gain after cleaning × current electricity price) - cleaning cost.

[0049] Calculate the expected net benefit B of cleaning delayed by 4 hours: (Expected power generation gain after cleaning × off-peak electricity price) - cleaning cost - (power generation loss due to pollution caused by the 4-hour delay).

[0050] Comparing A and B: If B > A, and B is the largest among all future selectable time periods, the system generates an instruction to "start cleaning 4 hours later (valley time period)" and sets a timer for triggering.

[0051] Step 5: Command Execution and Closed-Loop Correction. The execution module schedules the robot to complete the cleaning at a specified time. After completion, the sensing module collects power data indicating stable operation after cleaning.

[0052] Calculate the actual gain G_real: (Average power after cleaning - Average power before cleaning) × running time.

[0053] Search the historical records to obtain the estimated G_pred at the time of decision-making.

[0054] Calculate the deviation Δ = G_real - G_pred.

[0055] Update coefficients: Gradient descent is used, k0_new = k0_old + α * Δ (α is the learning rate, such as 0.001). The new k0 (i.e., k0_new) will be used for the next round of threshold calculation for this array, thereby correcting model bias. For example, if the actual gain is found to be lower than the prediction multiple times, k0 will be automatically lowered, making future decisions more conservative (the threshold is raised, making it more difficult to trigger cleaning).

[0056] In some embodiments of this application, a non-transitory computer-readable storage medium is provided, on which a program or instructions are stored. When the program or instructions are executed by a processor, they implement the steps of an adaptive cleaning decision method for a photovoltaic power station in a resource-unconstrained scenario as provided in any of the above embodiments. Therefore, the readable storage medium also includes all the beneficial effects of the method provided in any of the above embodiments. To avoid repetition, these will not be repeated here.

[0057] In some embodiments of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of an adaptive cleaning decision method for a photovoltaic power plant under resource-unconstrained scenarios. Therefore, this electronic device also includes all the beneficial effects of the methods provided in any of the above embodiments, and to avoid repetition, they will not be described again here.

[0058] Application example:

[0059] A year-long comparative verification was conducted at a 1.8MW distributed photovoltaic power station (containing 12 independent arrays) in Jiangsu Province. The method described in the embodiments of this application (Scheme C) was compared with traditional fixed-period cleaning (Scheme A) and a simple strategy based solely on a contamination rate threshold (Scheme B).

[0060] Key improvements are reflected in:

[0061] 1. Rainfall Game Effect: Option C successfully avoided the planned cleaning before 3 light rains (<10mm) and canceled the triggered machine cleaning after 2 heavy rains (>25mm), saving costs and avoiding the risk of mud stains.

[0062] 2. Optimized electricity pricing: More than 60% of cleaning instructions are optimized for execution during off-peak hours, significantly reducing cleaning energy costs.

[0063] 3. Closed-loop correction effect: After six months of system operation, the k0 coefficients of each array tend to stabilize, and the average error between the model's predicted gain and the actual gain decreases from the initial 15% to less than 5%, with the decision accuracy continuously improving.

[0064] The final annual comparison data is shown in the table below:

[0065]

[0066] Data shows that the solution in this application achieves higher power generation efficiency and net profit with fewer cleaning cycles, fully demonstrating its technological advancement and economic efficiency.

[0067] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An adaptive cleaning decision method for photovoltaic power plants under resource-unconstrained scenarios, characterized in that, Includes the following steps: S1: Using a single photovoltaic array and its dedicated cleaning robot as an independent decision-making unit, real-time pollution rate data, weather forecast data, time-of-use electricity price information, and equipment status data of the cleaning robot are collected in real time. S2: Calculate the dynamic contamination rate threshold for the independent decision-making unit, and compare the real-time contamination rate with the dynamic contamination rate threshold. If the real-time contamination rate does not exceed the dynamic contamination rate threshold, the decision is to remain on standby. S3: If the real-time pollution rate exceeds the dynamic pollution rate threshold, obtain the weather forecast within the future preset time window, calculate the net cleaning utility value of the forecast rainfall, and when the net cleaning utility value is less than the preset mud spot risk threshold, generate an instruction to immediately execute machine cleaning. S4: After determining that machine cleaning needs to be performed, combine the time-of-use electricity price curves for the next N hours to calculate the expected net benefit of immediate cleaning and the expected net benefit of cleaning during potential low-electricity-price periods in the future. Select the moment with the largest expected net benefit as the final operation start point and generate the corresponding cleaning instruction. S5: After the cleaning operation is completed, based on the actual operating data fed back by the dedicated cleaning robot, the actual energy efficiency improvement gain before and after the array cleaning is calculated, and the deviation between the actual energy efficiency improvement gain and the cleaning gain estimated when making the cleaning decision is calculated. Based on the deviation, the benchmark coefficient in the dynamic pollution rate threshold is iteratively updated through an online learning algorithm. The updated benchmark coefficient is used for the next decision cycle of the independent decision unit.

2. The adaptive cleaning decision method for photovoltaic power plants in resource-unconstrained scenarios according to claim 1, characterized in that, The dynamic pollution rate threshold θ S(t) The calculation formula is: Where k0 is the baseline coefficient updated through online learning; OC i P_avg(t) is the cost of a single cleaning operation for the i-th photovoltaic array; P_avg(t) is the predicted average electricity price over a future cleaning effect period; P_irated is the rated power of the photovoltaic array; k_season is the seasonal adjustment coefficient set based on the abundance of solar resources at the location of the photovoltaic array; k_weather is the weather adjustment coefficient set based on the continuity of microenvironment weather forecasts for the photovoltaic array.

3. The adaptive cleaning decision method for photovoltaic power plants in resource-unconstrained scenarios according to claim 2, characterized in that, The seasonal adjustment coefficient k_season takes a negative value in summer to encourage early cleaning and a positive value in winter to increase the threshold.

4. The adaptive cleaning decision method for photovoltaic power plants in resource-unconstrained scenarios according to claim 2, characterized in that, The weather adjustment coefficient k_weather is set to a negative value to lower the threshold when consecutive sunny days are forecast, and to a positive value to raise the threshold when intermittent dust or rain is forecast.

5. An adaptive cleaning decision method for photovoltaic power plants in resource-unconstrained scenarios according to any one of claims 2 to 4, characterized in that, In step S3: When the forecasted rainfall is below the first threshold, the net cleaning utility value is determined to be negative; When the forecasted rainfall is higher than the second threshold, the net cleaning utility value is determined to be positive. When the forecast rainfall is between the first and second thresholds, the probability of mud spot effect is assessed and the net cleaning utility value is calculated based on the historical data model.

6. The adaptive cleaning decision method for photovoltaic power plants in resource-unconstrained scenarios according to claim 5, characterized in that, The future N hours mentioned in step S4 are 24 hours, and the formula for calculating the expected net income is: Expected net income = (expected power generation gain after cleaning × electricity price for the corresponding period) - cleaning cost - power generation loss caused by pollution during the delay period.

7. The adaptive cleaning decision method for photovoltaic power plants in resource-unconstrained scenarios according to claim 1, characterized in that, The meteorological forecast data is gridded meteorological data containing rainfall forecasts within 72 hours, and the time-of-use electricity price information is obtained in real time through the power grid marketing system interface.

8. An adaptive cleaning decision system for photovoltaic power plants under resource-unconstrained scenarios, characterized in that, include: The sensing module is used to collect pollution rate data, weather forecast data, time-of-use electricity price information, and equipment status data of the corresponding dedicated cleaning robot in real time for each independent decision-making unit, on a per-photovoltaic array basis. The decision module, which is communicatively connected to the perception module and deployed on an edge computing node close to the photovoltaic array, is used to independently execute steps S2-S5 of the method according to any one of claims 1-7 for each independent decision unit to generate cleaning instructions. The execution module, which is communicatively connected to the decision module, is used to receive the cleaning instructions and dispatch the corresponding dedicated cleaning robot to perform the cleaning operation, and to feed back the operation results to the perception module to form a closed loop.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images