Photovoltaic power station dynamic cleaning decision method and system based on digital twinning and multi-objective optimization

Abstract

The present application relates to the technical field of digital twinning, in particular to a photovoltaic power station dynamic cleaning decision method and system based on digital twinning and multi-objective optimization, comprising the following steps: collecting geographic information data, component arrangement topology data and electrical connection relationship data of the photovoltaic power station, and constructing a photovoltaic power station digital twinning model according to the geographic information data, the component arrangement topology data and the electrical connection relationship data.In the present application, a multi-objective genetic algorithm is used to generate a dynamic scheme containing a specific time window and path planning, which can guide the equipment to execute, and through the collection of actual execution power generation gain and cost data to calculate the reward value and update the decision weight by using the reinforcement learning algorithm, a perception decision execution feedback closed-loop optimization mechanism is constructed, so that the system has the ability of continuous self-learning evolution, and the strategy preference is dynamically adjusted according to the actual operation effect, thereby improving the operation and maintenance economy and the decision intelligence level of the photovoltaic power station.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, and in particular to a dynamic cleaning decision-making method and system for photovoltaic power plants based on digital twins and multi-objective optimization. Background Technology

[0002] Digital twin technology is a simulation process that fully utilizes data such as physical models, sensor updates, and operational history to integrate multiple disciplines, multiple physical quantities, multiple scales, and multiple probabilities. It completes the mapping in virtual space, thereby reflecting the entire life cycle of the corresponding physical equipment.

[0003] Existing technologies often overlook the continuous impact of meteorological factors on cleaning effectiveness. For example, performing mechanical cleaning operations just before natural rainfall results in unnecessary waste of water resources, equipment power, and mechanical wear. Furthermore, natural rainfall itself has a cleaning function, rendering human intervention worthless. Additionally, relying solely on the physical thickness of accumulated ash while ignoring the direct impact of time-of-use electricity price fluctuations on output value can lead to high cleaning costs during periods of low electricity prices, resulting in an imbalance between input and output, and even situations where cleaning costs exceed the power generation gains, leading to negative returns. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies that often neglect the continuous interference of meteorological factors on the cleaning effect in actual operation, such as performing mechanical cleaning operations just before natural precipitation, resulting in unnecessary consumption of water resources, equipment power, and mechanical wear. The invention proposes a dynamic cleaning decision-making method and system for photovoltaic power plants based on digital twins and multi-objective optimization.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a dynamic cleaning decision-making method for photovoltaic power plants based on digital twins and multi-objective optimization, comprising the following steps:

[0006] Collect geographic information data, component layout topology data, and electrical connection relationship data of photovoltaic power plants, and construct a digital twin model of the photovoltaic power plant based on the geographic information data, component layout topology data, and electrical connection relationship data;

[0007] Real-time irradiance data and component surface dust monitoring data are acquired, and the real-time irradiance data and component surface dust monitoring data are input into the digital twin model of the photovoltaic power station to simulate the real-time dust accumulation distribution state of the photovoltaic components. The theoretical power generation loss value is calculated based on the real-time dust accumulation distribution state of the photovoltaic components.

[0008] Acquire short-term weather forecast data, time-of-use electricity price data, and cleaning equipment status data. Construct a multi-objective cleaning decision optimization model based on the theoretical power generation loss value, the short-term weather forecast data, the time-of-use electricity price data, and the cleaning equipment status data. The multi-objective cleaning decision optimization model is configured with decision weight parameters, and the optimization objectives are to maximize power generation revenue and minimize cleaning costs.

[0009] The multi-objective cleaning decision optimization model is solved using a multi-objective genetic algorithm to generate a dynamic cleaning decision scheme, wherein the dynamic cleaning decision scheme includes a cleaning time window sequence and a cleaning path planning instruction, and the cleaning operation is executed according to the dynamic cleaning decision scheme;

[0010] The actual power generation gain data and actual cleaning cost data after the cleaning operation are collected. The decision reward value is calculated based on the actual power generation gain data and the actual cleaning cost data. The decision weight parameters of the multi-objective cleaning decision optimization model are updated based on the decision reward value using a reinforcement learning algorithm.

[0011] Preferably, the steps for constructing a digital twin model of a photovoltaic power plant include:

[0012] The geographic information data is analyzed for topographic features to obtain a three-dimensional terrain model;

[0013] Based on the component layout topology data, the spatial positions of the photovoltaic components are mapped in the three-dimensional terrain model to generate a virtual component array model;

[0014] Based on the electrical connection relationship data, the electrical logic connection of the virtual component array model is established to obtain the digital twin model of the photovoltaic power station, wherein the digital twin model of the photovoltaic power station includes the physical attribute parameters and photoelectric conversion efficiency parameters of the photovoltaic components.

[0015] Preferably, the steps for calculating the theoretical power generation loss value include:

[0016] Based on the real-time dust accumulation distribution of the photovoltaic modules, the dust accumulation shading coefficient of each photovoltaic subarray is calculated.

[0017] The theoretical clean power generation capacity under dust-free conditions was simulated using the aforementioned digital twin model of the photovoltaic power station;

[0018] The theoretical clean power generation is corrected based on the ash accumulation shielding coefficient to obtain the power generation under ash accumulation conditions.

[0019] The difference between the theoretical clean power generation and the power generation under the ash accumulation state is calculated to obtain the real-time power loss value. The real-time power loss value is then integrated with a preset prediction time period to obtain the theoretical power generation loss value for each photovoltaic subarray.

[0020] Preferably, the steps for constructing a multi-objective cleaning decision optimization model include:

[0021] Based on the time-of-use electricity price data, the decision-making time period is determined. The on-grid electricity price coefficient;

[0022] The water consumption cost and equipment wear and depreciation cost for a single cleaning operation are determined based on the status data of the cleaning equipment.

[0023] Constructing the power generation revenue function and cleaning cost function And set the optimization objective as maximizing And minimize Power generation revenue function and cleaning cost function The function expressions are as follows:

[0024]

[0025]

[0026] in, This indicates the total number of photovoltaic subarrays. Indicates the first The theoretical power generation loss value of each photovoltaic sub-array within the predicted time period. Indicates the first The power generation recovery rate of each photovoltaic subarray after cleaning Indicates the decision-making period The aforementioned on-grid electricity price coefficient, This represents the unit price of water resources per unit volume. This indicates the total water usage during the cleaning process. This indicates the wear and tear cost per unit time for cleaning equipment. This indicates the total working time of the cleaning equipment.

[0027] Preferably, the steps of constructing a multi-objective cleaning decision optimization model also include:

[0028] Extract precipitation probability values ​​and wind speed prediction values ​​from the short-term weather forecast data;

[0029] Set precipitation probability thresholds and wind speed safety thresholds;

[0030] A first constraint condition is established, which is that when the precipitation probability value is greater than the precipitation probability threshold, the cleaning time window sequence is forcibly set to a non-working state.

[0031] A second constraint is constructed, which is to stop generating the cleaning path planning instruction when the predicted wind speed value is greater than the wind speed safety threshold.

[0032] The first constraint and the second constraint are added to the multi-objective cleaning decision optimization model.

[0033] Preferably, the step of acquiring dust monitoring data on the surface of the component includes:

[0034] Collect multispectral image data of the photovoltaic module surface;

[0035] Texture features are extracted from the multispectral image data using a convolutional neural network model to obtain dust type classification results, wherein the dust type classification results include sand dust type, pollen type and oil stain type;

[0036] Determine the corresponding dust adhesion coefficient based on the dust type classification results;

[0037] The dust adhesion coefficient is used as part of the dust monitoring data on the component surface to correct the dust transmittance parameter in the digital twin model of the photovoltaic power station.

[0038] Preferably, the step of generating a dynamic cleaning decision plan further includes:

[0039] Match the corresponding cleaning medium parameters according to the dust type classification result. When the dust type classification result is the oil stain type, increase the cleaning fluid concentration parameter.

[0040] Adjust the brush head rotation speed and travel speed parameters of the cleaning device according to the dust adhesion coefficient;

[0041] The cleaning medium parameters, the brush head rotation speed parameters, and the travel speed parameters are encapsulated into the dynamic cleaning decision scheme.

[0042] Preferably, the step of generating the cleaning path planning instruction includes:

[0043] Based on the theoretical power generation loss value, the multiple sub-array areas within the photovoltaic power station are sorted to obtain a power generation benefit priority sequence;

[0044] Obtain the remaining battery power and current location coordinates from the status data of the cleaning device;

[0045] When solving the cleaning path planning instructions, the ant colony algorithm is used to search for the optimal traversal path in the power generation benefit priority sequence, wherein the optimal traversal path is optimized to cover the high power generation benefit area and have the shortest total moving distance.

[0046] The optimal traversal path is converted into the cleaning path planning instruction.

[0047] Preferably, the step of updating the decision weight parameters includes:

[0048] The difference between the actual power generation gain data and the actual cleaning cost data is calculated to obtain the actual net profit value;

[0049] The expected net profit is calculated based on the power generation revenue and cleaning cost predicted by the multi-objective cleaning decision optimization model.

[0050] Calculate the deviation rate between the actual net income value and the expected net income value;

[0051] Construct a reinforcement learning reward function, and output a positive reward value when the deviation rate is less than a preset deviation threshold;

[0052] The decision weight parameters in the multi-objective cleaning decision optimization model are updated using the positive reward value, wherein the decision weight parameters include power generation revenue weight and cleaning cost weight;

[0053] The steps for solving the multi-objective cleaning decision optimization model using a multi-objective genetic algorithm include:

[0054] Initialize a cleaning strategy population, wherein the cleaning strategy population includes several randomly generated combinations of cleaning time and path;

[0055] According to the power generation revenue function and the cleaning cost function Calculate the non-dominated ranking rank and crowding distance of each individual in the population of the cleaning strategy;

[0056] The cleaning strategy population is subjected to selection, crossover, and mutation operations to generate offspring populations;

[0057] Merge the sweeping strategy population and the offspring population to select the Pareto optimal front solution set;

[0058] Obtain the current decision weight parameters of the multi-objective cleaning decision optimization model, and denote them as the power generation revenue weight. and cleaning cost weighting ;

[0059] Based on the power generation revenue weight and the cleaning cost weight Calculate the comprehensive decision score for each solution in the Pareto optimal frontier solution set. The formula for calculating the comprehensive decision score is as follows: The solution with the highest comprehensive decision score is selected as the dynamic cleaning decision scheme.

[0060] The present invention also provides a system comprising:

[0061] Digital twin modeling module: used to collect geographic information data, component layout topology data and electrical connection relationship data of photovoltaic power station, and to construct a digital twin model of photovoltaic power station based on the geographic information data, component layout topology data and electrical connection relationship data;

[0062] Dust accumulation simulation and power generation loss assessment module: used to acquire real-time irradiance data and dust monitoring data on the surface of the components, input the real-time irradiance data and the dust monitoring data on the surface of the components into the digital twin model of the photovoltaic power station, simulate the real-time dust accumulation distribution state of the photovoltaic components, and calculate the theoretical power generation loss value based on the real-time dust accumulation distribution state of the photovoltaic components.

[0063] Multi-objective cleaning optimization modeling module: used to acquire short-term weather forecast data, time-of-use electricity price data and cleaning equipment status data, and to construct a multi-objective cleaning decision optimization model based on the theoretical power generation loss value, the short-term weather forecast data, the time-of-use electricity price data and the cleaning equipment status data. The multi-objective cleaning decision optimization model is configured with decision weight parameters and takes maximizing power generation revenue and minimizing cleaning costs as optimization objectives.

[0064] Dynamic cleaning decision generation module: used to solve the multi-objective cleaning decision optimization model using a multi-objective genetic algorithm, generate a dynamic cleaning decision scheme, wherein the dynamic cleaning decision scheme includes a cleaning time window sequence and a cleaning path planning instruction, and executes the cleaning operation according to the dynamic cleaning decision scheme;

[0065] Model Adaptive Optimization Module: Used to collect actual power generation gain data and actual cleaning cost data after the cleaning operation is performed, calculate the decision reward value based on the actual power generation gain data and the actual cleaning cost data, and use a reinforcement learning algorithm to update the decision weight parameters of the multi-objective cleaning decision optimization model based on the decision reward value.

[0066] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0067] This invention constructs a digital twin model by collecting geographical information, component layout topology, and electrical connection data of photovoltaic power plants. This model accurately maps all elements of the physical power plant to a virtual space. By combining real-time irradiance and dust monitoring data for simulation, it can reconstruct the real-time dust accumulation distribution of photovoltaic modules and quantify the theoretical power generation loss, ensuring that cleaning decisions are based on energy efficiency data rather than fuzzy empirical estimations. A multi-objective optimization model is constructed by introducing short-term weather forecasts, time-of-use electricity prices, and cleaning equipment status data. This model comprehensively considers the impact of weather conditions on the duration of cleaning effectiveness and the impact of electricity price fluctuations on power generation revenue, seeking the optimal solution between maximizing power generation revenue and minimizing cleaning costs. This avoids ineffective cleaning before rainfall or high-cost operations during periods of low electricity prices. A multi-objective genetic algorithm generates dynamic solutions including specific time windows and path planning, guiding equipment execution. By collecting power generation gain and cost data after actual execution, reward values ​​are calculated, and reinforcement learning algorithms are used to update decision weights, constructing a closed-loop optimization mechanism for perception, decision-making, and execution feedback. This enables the system to continuously learn and evolve, dynamically adjusting strategy preferences based on actual operating results, thereby improving the economic efficiency and intelligent decision-making level of photovoltaic power plant operation and maintenance. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0070] Please see Figure 1 This invention provides a technical solution: a dynamic cleaning decision-making method for photovoltaic power plants based on digital twins and multi-objective optimization, comprising the following steps:

[0071] Collect geographic information data, component layout topology data, and electrical connection relationship data of photovoltaic power plants, and construct a digital twin model of the photovoltaic power plant based on the geographic information data, component layout topology data, and electrical connection relationship data;

[0072] Real-time irradiance data and component surface dust monitoring data are acquired, and the real-time irradiance data and component surface dust monitoring data are input into the digital twin model of the photovoltaic power station to simulate the real-time dust accumulation distribution state of the photovoltaic components. The theoretical power generation loss value is calculated based on the real-time dust accumulation distribution state of the photovoltaic components.

[0073] Acquire short-term weather forecast data, time-of-use electricity price data, and cleaning equipment status data. Construct a multi-objective cleaning decision optimization model based on the theoretical power generation loss value, the short-term weather forecast data, the time-of-use electricity price data, and the cleaning equipment status data. The multi-objective cleaning decision optimization model is configured with decision weight parameters, and the optimization objectives are to maximize power generation revenue and minimize cleaning costs.

[0074] A multi-objective genetic algorithm is used to solve the multi-objective cleaning decision optimization model and generate a dynamic cleaning decision scheme. The dynamic cleaning decision scheme includes a cleaning time window sequence and a cleaning path planning instruction. The cleaning operation is executed according to the dynamic cleaning decision scheme.

[0075] The system collects actual power generation gain data and actual cleaning cost data after the cleaning operation is performed. Based on the actual power generation gain data and actual cleaning cost data, it calculates the decision reward value and uses a reinforcement learning algorithm to update the decision weight parameters of the multi-objective cleaning decision optimization model according to the decision reward value.

[0076] The steps for constructing a digital twin model of a photovoltaic power station include: performing topographic analysis on geographic information data to obtain a three-dimensional terrain model; mapping the spatial location of photovoltaic modules in the three-dimensional terrain model based on the module layout topology data to generate a virtual module array model; and establishing the electrical logic connection of the virtual module array model based on electrical connection relationship data to obtain a digital twin model of the photovoltaic power station. The digital twin model of the photovoltaic power station includes the physical attribute parameters and photoelectric conversion efficiency parameters of the photovoltaic modules.

[0077] Specifically, topographic analysis is performed on the geographic information data. Digital elevation model (DEM) data and orthophoto data are read from the geographic information data. An irregular triangular mesh (ITM) construction algorithm is used to grid the DEM data. The elevation sampling interval is set to a range of 5 to 10 meters. Elevation and coordinate values ​​of topographic feature points are extracted. Topographic undulation features are smoothed using spline interpolation to construct a 3D topographic model consistent with the actual terrain. The spatial location of photovoltaic (PV) modules is mapped into the 3D topographic model based on the module layout topology data. CAD engineering drawings from the module layout topology data are read, and the corner coordinates and tilt azimuth parameters of the PV array in the drawings are analyzed. A coordinate transformation matrix is ​​used to convert the drawing coordinate system to the geographic coordinate system of the 3D topographic model. The 3D model of each PV module is instantiated and anchored to its corresponding latitude, longitude, and altitude coordinates, generating a virtual... The module array model establishes the electrical logic connections of the virtual module array model based on electrical connection relationship data. It traverses the DC combiner box wiring table and string inverter configuration table within the electrical connection relationship data, constructing a directed acyclic graph data structure to represent current flow and series-parallel relationships according to the hierarchical order from photovoltaic cells, photovoltaic modules, photovoltaic strings, combiner boxes to inverters. A unique electrical identifier is assigned to each node, resulting in a digital twin model of the photovoltaic power station. This digital twin model includes the physical attribute parameters and photoelectric conversion efficiency parameters of the photovoltaic modules. Physical attribute parameters include module size, glass transmittance, and surface roughness. Photoelectric conversion efficiency parameters include the open-circuit voltage temperature coefficient, short-circuit current temperature coefficient, and maximum power point voltage temperature coefficient. The temperature coefficients are entered based on actual measured data from the module's factory, for example, the open-circuit voltage temperature coefficient is set to -0.28% per degree Celsius.

[0078] The steps for calculating the theoretical power generation loss value include: calculating the shading coefficient of each photovoltaic subarray based on the real-time ash distribution of the photovoltaic modules; simulating the theoretical clean power generation in a ash-free state using a digital twin model of the photovoltaic power station; correcting the theoretical clean power generation based on the shading coefficient to obtain the power generation in the ash-accumulated state; calculating the difference between the theoretical clean power generation and the power generation in the ash-accumulated state to obtain the real-time power loss value; and integrating the real-time power loss value with a preset prediction time period to obtain the theoretical power generation loss value for each photovoltaic subarray.

[0079] Specifically, based on the real-time dust accumulation distribution of the photovoltaic modules, the dust accumulation shading coefficient of each photovoltaic subarray is calculated. The pixel-level dust accumulation thickness in the real-time dust accumulation distribution of the photovoltaic modules is mapped to a transmittance attenuation matrix. For modules containing... The first series component For each photovoltaic subarray, the weighted average transmittance of all components within the subarray is calculated as the dust accumulation and shading coefficient. If a heavily dusty area exists within the subarray, the transmittance of the series circuit is limited to the lowest transmittance value of the component based on the barrel effect. A digital twin model of the photovoltaic power station is used to simulate the theoretical clean power generation under dust-free conditions. The photoelectric conversion efficiency parameters and real-time meteorological irradiance data from the photovoltaic power station's digital twin model are used. A single-diode mathematical model is employed to calculate the output characteristics of the photovoltaic module. Input parameters include photocurrent, reverse saturation current, diode ideality factor, series resistance, and parallel resistance. The current-voltage equations are solved using the Newton-Raphson iterative method to obtain the standard test conditions. The output power is calculated, and the power is corrected according to the real-time component temperature to obtain the theoretical clean power generation. The theoretical clean power generation is corrected according to the dust accumulation shading coefficient to obtain the power generation in the dust accumulation state. The theoretical clean power generation is multiplied by the corresponding transmittance retention coefficient, which is 1 minus the dust accumulation shading coefficient. The difference between the theoretical clean power generation and the power generation in the dust accumulation state is calculated to obtain the real-time power loss value. The real-time power loss value is integrated in combination with the preset prediction time period. The prediction time period is set to the next 24 hours, and the integration step is 15 minutes. The real-time power loss value in each time step is accumulated to obtain the theoretical power generation loss value of each photovoltaic subarray.

[0080] The steps for constructing a multi-objective cleaning decision optimization model include: determining the decision time period based on time-of-use electricity price data. The on-grid electricity price coefficient is used to determine the water consumption cost and equipment wear and depreciation cost of a single cleaning operation based on the status data of the cleaning equipment, and a power generation revenue function is constructed. and cleaning cost function And set the optimization goal to maximize And minimize Power generation revenue function and cleaning cost function The function expressions are as follows: , ,in, This indicates the total number of photovoltaic subarrays. Indicates the first The theoretical power generation loss of each photovoltaic sub-array within the predicted time period. Indicates the first The power generation recovery rate of each photovoltaic subarray after cleaning Indicates the decision-making period The on-grid electricity price coefficient, This represents the unit price of water resources per unit volume. This indicates the total water usage during the cleaning process. This indicates the wear and tear cost per unit time for cleaning equipment. This indicates the total working time of the cleaning equipment.

[0081] Specifically, the decision-making time period is determined based on time-of-use electricity price data. The grid connection tariff coefficient is determined by retrieving the latest industrial and commercial photovoltaic grid connection tariff table published by the local power grid company, analyzing the peak, flat, and valley time period division rules, dividing a 24-hour day into 96 15-minute time nodes, and matching a corresponding electricity price value to each time node. For example, the electricity price coefficient is set at 1.2 yuan per kilowatt-hour during the peak period from 10:00 to 14:00, and at 0.3 yuan per kilowatt-hour during the valley period from 00:00 to 06:00. Based on the cleaning equipment status data, the water consumption cost and equipment wear and depreciation cost of a single cleaning operation are determined. The water tank capacity parameters and nozzle flow parameters of the cleaning robot are read, and the water cost for a single complete operation is calculated in combination with the local industrial water price. The rated life of the robot's motor and the replacement cycle of key components are read, and the wear and tear amortization cost per unit operating time is calculated using the straight-line depreciation method. Finally, a power generation revenue function is constructed. and cleaning cost function And set the optimization goal to maximize And minimize Power generation revenue function and cleaning cost function The function expressions are as follows: , ,in, This indicates the total number of photovoltaic subarrays. Indicates the first The theoretical power generation loss of each photovoltaic sub-array within the predicted time period. Indicates the first The power generation recovery rate of each photovoltaic subarray after cleaning is calculated based on historical cleaning results and ranges from 0.95 to 0.99. Indicates the decision-making period The on-grid electricity price coefficient, This represents the unit price of water resources per unit volume. This indicates the total water usage during the cleaning process. This indicates the wear and tear cost per unit time for cleaning equipment. This indicates the total working time of the cleaning equipment.

[0082] The steps for constructing a multi-objective cleaning decision optimization model also include: extracting precipitation probability values ​​and wind speed prediction values ​​from short-term weather forecast data, setting precipitation probability thresholds and wind speed safety thresholds, constructing a first constraint condition, which is to force the cleaning time window sequence to be in a non-working state when the precipitation probability value is greater than the precipitation probability threshold, constructing a second constraint condition, which is to stop generating cleaning path planning instructions when the wind speed prediction value is greater than the wind speed safety threshold, and adding the first constraint condition and the second constraint condition to the multi-objective cleaning decision optimization model.

[0083] Specifically, precipitation probability and wind speed forecasts are extracted from short-term weather forecast data. The JSON data packet returned by the weather service interface is parsed to read hourly precipitation probability percentage data and wind speed values ​​at a height of 10 meters for the next 72 hours. Precipitation probability thresholds and wind speed safety thresholds are set. Based on historical invalid cleaning case statistics, the precipitation probability threshold is set to 40%. Based on the cleaning robot's wind resistance test report, the wind speed safety threshold is set to 10 meters per second. A first constraint is constructed: when the precipitation probability value is greater than the precipitation probability threshold, the cleaning time window sequence is forcibly set to a non-working state. This involves introducing Boolean variables into the optimization model. When the probability of precipitation exceeds 40% during a certain period, let the decision variable corresponding to that period be... To prevent cleaning tasks from being scheduled during this period, a second constraint is established: when the predicted wind speed exceeds the wind speed safety threshold, the generation of cleaning path planning instructions is stopped. This means that a penalty function is introduced into the optimization model. When the wind speed exceeds 10 meters per second, the objective function value is set to infinity to eliminate the infeasible solution. The first and second constraints are then added to the multi-objective cleaning decision optimization model.

[0084] The steps for obtaining dust monitoring data on the surface of photovoltaic modules include: collecting multispectral image data of the photovoltaic module surface, extracting texture features from the multispectral image data using a convolutional neural network model, obtaining dust type classification results, including sand dust, pollen, and oil stains, determining the corresponding dust adhesion coefficient based on the dust type classification results, and using the dust adhesion coefficient as part of the dust monitoring data on the module surface to correct the dust transmittance parameter in the digital twin model of the photovoltaic power station.

[0085] Specifically, multispectral image data of the photovoltaic module surface is collected. An inspection drone equipped with a multispectral camera is controlled to hover at a height of 3 to 5 meters above the module surface to capture images, obtaining image data in five bands: red, green, blue, near-infrared, and red-edge. A convolutional neural network (CNN) model is used to extract texture features from the multispectral image data. A deep CNN based on the ResNet-50 architecture is constructed. The CNN model includes an input layer, multiple residual convolutional blocks, a global average pooling layer, and a fully connected output layer. The input layer receives five-channel tensor data of size 256 x 256 pixels and uses 3 x 3 convolutional kernels for feature extraction. After processing by a batch normalization layer and the ReLU activation function, the data is input to the residual blocks. During the training phase, the data is used... Supervised training was performed on a labeled dust sample dataset. The network weights were optimized using the cross-entropy loss function. The model parameters were iteratively updated using the backpropagation algorithm until the validation set accuracy exceeded 95%, resulting in dust type classification results. These results included sand, pollen, and oil types. Based on the dust type classification results, the corresponding dust adhesion coefficients were determined, and a mapping table between dust type and adhesion coefficient was established. For example, the adhesion coefficient for sand was set to 0.2, for pollen to 0.5, and for oil to 0.8. The dust adhesion coefficients were used as part of the dust monitoring data on the component surface to correct the dust transmittance parameter in the digital twin model of the photovoltaic power station.

[0086] The steps for generating a dynamic cleaning decision scheme also include: matching the corresponding cleaning medium parameters according to the dust type classification results; when the dust type classification result is oil, adding the cleaning fluid concentration parameter; adjusting the brush head rotation speed parameter and travel speed parameter of the cleaning equipment according to the dust adhesion coefficient; and encapsulating the cleaning medium parameters, brush head rotation speed parameters, and travel speed parameters into the dynamic cleaning decision scheme.

[0087] Specifically, based on the dust type classification result, the corresponding cleaning medium parameters are matched, and the preset cleaning process reference table is consulted. If the dust type classification result is sand and dust, the waterless dry brush mode is selected; if the result is pollen, the water rinsing mode is selected. When the dust type classification result is oil, the cleaning fluid concentration parameter is increased, setting the mixing ratio of cleaning fluid to water to 1:20, and this ratio parameter is written into the decision instruction package. The brush head rotation speed and travel speed parameters of the cleaning equipment are adjusted according to the dust adhesion coefficient, utilizing linear... The regression model calculates specific parameter values, setting the baseline brush head rotation speed to 200 revolutions per minute and the baseline travel speed to 0.5 meters per second. The calculation formula is: the set rotation speed equals the baseline rotation speed plus the adhesion coefficient multiplied by 100, and the set travel speed equals the baseline travel speed minus the adhesion coefficient multiplied by 0.2. For example, when the adhesion coefficient is 0.8, the calculated brush head rotation speed is 280 revolutions per minute and the travel speed is 0.34 meters per second. The cleaning medium parameters, brush head rotation speed parameters, and travel speed parameters are encapsulated into the dynamic cleaning decision scheme.

[0088] The steps for generating cleaning path planning instructions include: sorting multiple sub-array areas within the photovoltaic power station according to the theoretical power generation loss value to obtain a power generation benefit priority sequence; obtaining the remaining power data and current position coordinates from the cleaning equipment status data; and using the ant colony algorithm to search for the optimal traversal path in the power generation benefit priority sequence when solving for the cleaning path planning instructions. The optimal traversal path aims to cover high power generation benefit areas and minimize the total moving distance. The optimal traversal path is then converted into a cleaning path planning instruction.

[0089] Specifically, based on the theoretical power generation loss value, multiple sub-array areas within the photovoltaic power station are sorted. All sub-arrays are arranged from largest to smallest according to the calculated loss value. The top few sub-arrays with a cumulative loss ratio reaching 80% are selected as priority cleaning targets, resulting in a power generation benefit priority sequence. The remaining power data and current location coordinates from the cleaning equipment status data are obtained. The state of charge percentage value and differential GPS positioning data uploaded by the battery management system are read. When solving the cleaning path planning instructions, the ant colony algorithm is used to search for the optimal traversal path in the power generation benefit priority sequence. A graph model with the sub-array center coordinates as nodes is constructed. A group of virtual ants is initialized and distributed at the current position of the cleaning equipment. A pheromone matrix and a heuristic function are set. The value of the heuristic function is inversely proportional to the Euclidean distance between nodes and directly proportional to the power generation benefit priority of the node. Each ant selects the next node to visit according to the probability transition rule. The probability transition formula is: ,in Indicates the first Only ants from the node Transfer to node The probability, The pheromone concentration along the path, For heuristic information, and These are the pheromone heuristic factor and the expectation heuristic factor, respectively. After completing one traversal, additional pheromone increments are applied to paths with short path lengths and many high-priority nodes. The traversal converges through multiple iterations. The optimal traversal path aims to cover high-power-efficiency areas and minimize the total travel distance. The optimal traversal path is then converted into a sweeping path planning instruction.

[0090] The steps for updating decision weight parameters include: calculating the difference between actual power generation gain data and actual cleaning cost data to obtain the actual net benefit value; calculating the expected net benefit value based on the power generation benefit value and cleaning cost value predicted by the multi-objective cleaning decision optimization model; calculating the deviation rate between the actual net benefit value and the expected net benefit value; constructing a reinforcement learning reward function; outputting a positive reward value when the deviation rate is less than a preset deviation threshold; and updating the decision weight parameters in the multi-objective cleaning decision optimization model using the positive reward value, wherein the decision weight parameters include power generation benefit weight and cleaning cost weight. The steps for solving the multi-objective cleaning decision optimization model using a multi-objective genetic algorithm include: initializing the cleaning strategy population, wherein the cleaning strategy population contains several randomly generated combinations of cleaning time and path; and calculating the actual net benefit value based on the power generation benefit function. and cleaning cost function The non-dominated ranking rank and crowding distance of each individual in the cleaning strategy population are calculated. Selection, crossover, and mutation operations are performed on the cleaning strategy population to generate offspring populations. The cleaning strategy populations and offspring populations are merged, and the Pareto optimal frontier solution set is selected. The current decision weight parameters of the multi-objective cleaning decision optimization model are obtained, denoted as the power generation revenue weight. and cleaning cost weighting According to the power generation revenue weight and cleaning cost weighting Calculate the comprehensive decision score for each solution in the Pareto optimal frontier solution set. The formula for calculating the comprehensive decision score is: The solution with the highest comprehensive decision score is selected as the dynamic cleaning decision scheme.

[0091] Specifically, the steps for updating decision weight parameters include: calculating the difference between actual power generation gain data and actual cleaning cost data; reading the actual power generation increment recorded by the electricity meter after cleaning and the actual expenditure cost calculated by the financial system, subtracting the two to obtain the actual net benefit value; calculating the expected net benefit value based on the power generation benefit value and cleaning cost value predicted by the multi-objective cleaning decision optimization model; calculating the deviation rate between the actual net benefit value and the expected net benefit value, using the formula: the deviation rate equals the absolute value (actual net benefit value minus expected net benefit value) divided by the expected net benefit value; constructing a reinforcement learning reward function, setting a positive reward value when the deviation rate is less than a preset deviation threshold (e.g., 5%). Otherwise, a negative reward value will be given. The model outputs a positive reward value, which is used to update the decision weight parameters in the multi-objective cleaning decision optimization model. The weights are adjusted using a policy gradient algorithm. The decision weight parameters include power generation revenue weights and cleaning cost weights. The steps for solving the multi-objective cleaning decision optimization model using a multi-objective genetic algorithm include: initializing the cleaning strategy population, which contains several randomly generated combinations of cleaning time and path, setting the population size to 100, and encoding each individual as a one-dimensional array containing the cleaning start time and cleaning order. The model is then adjusted based on the power generation revenue function. and cleaning cost function The non-dominated ranking rank and crowding distance of each individual in the cleaning strategy population are calculated. The two objective function values ​​are compared. If individual A is not inferior to individual B in all objectives and is superior to individual B in at least one objective, then A is said to dominate B. Ranks are assigned based on the dominance relationship, and crowding distance is calculated within the same rank to maintain diversity. Selection, crossover, and mutation operations are performed on the cleaning strategy population. The tournament selection method is used to select the parent generation, the simulated binary crossover operator is used for crossover, and the polynomial mutation operator is used for mutation to generate the offspring population. The cleaning strategy population and the offspring population are merged, and the Pareto optimal frontier solution set is selected. The current decision weight parameters of the multi-objective cleaning decision optimization model are obtained and denoted as the power generation revenue weight. and cleaning cost weighting According to the power generation revenue weight and cleaning cost weighting Calculate the comprehensive decision score for each solution in the Pareto optimal frontier solution set. The formula for calculating the comprehensive decision score is: The solution with the highest comprehensive decision score is selected as the dynamic cleaning decision scheme.

[0092] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A dynamic cleaning decision-making method for photovoltaic power plants based on digital twins and multi-objective optimization, characterized in that, Includes the following steps: Collect geographic information data, component layout topology data, and electrical connection relationship data of photovoltaic power plants, and construct a digital twin model of the photovoltaic power plant based on the geographic information data, component layout topology data, and electrical connection relationship data; Real-time irradiance data and component surface dust monitoring data are acquired, and the real-time irradiance data and component surface dust monitoring data are input into the digital twin model of the photovoltaic power station to simulate the real-time dust accumulation distribution state of the photovoltaic components. The theoretical power generation loss value is calculated based on the real-time dust accumulation distribution state of the photovoltaic components. Acquire short-term weather forecast data, time-of-use electricity price data, and cleaning equipment status data. Construct a multi-objective cleaning decision optimization model based on the theoretical power generation loss value, the short-term weather forecast data, the time-of-use electricity price data, and the cleaning equipment status data. The multi-objective cleaning decision optimization model is configured with decision weight parameters, and the optimization objectives are to maximize power generation revenue and minimize cleaning costs. The multi-objective cleaning decision optimization model is solved using a multi-objective genetic algorithm to generate a dynamic cleaning decision scheme, wherein the dynamic cleaning decision scheme includes a cleaning time window sequence and a cleaning path planning instruction, and the cleaning operation is executed according to the dynamic cleaning decision scheme; The actual power generation gain data and actual cleaning cost data after the cleaning operation are collected. The decision reward value is calculated based on the actual power generation gain data and the actual cleaning cost data. The decision weight parameters of the multi-objective cleaning decision optimization model are updated based on the decision reward value using a reinforcement learning algorithm.

2. The photovoltaic power plant dynamic cleaning decision-making method based on digital twin and multi-objective optimization according to claim 1, characterized in that, The steps for constructing a digital twin model of a photovoltaic power plant include: The geographic information data is analyzed for topographic features to obtain a three-dimensional terrain model; Based on the component layout topology data, the spatial positions of the photovoltaic components are mapped in the three-dimensional terrain model to generate a virtual component array model; Based on the electrical connection relationship data, the electrical logic connection of the virtual component array model is established to obtain the digital twin model of the photovoltaic power station, wherein the digital twin model of the photovoltaic power station includes the physical attribute parameters and photoelectric conversion efficiency parameters of the photovoltaic components.

3. The dynamic cleaning decision-making method for photovoltaic power plants based on digital twins and multi-objective optimization according to claim 1, characterized in that, The steps for calculating the theoretical power generation loss include: Based on the real-time dust accumulation distribution of the photovoltaic modules, the dust accumulation shading coefficient of each photovoltaic subarray is calculated. The theoretical clean power generation capacity under dust-free conditions was simulated using the aforementioned digital twin model of the photovoltaic power station; The theoretical clean power generation is corrected based on the ash accumulation shielding coefficient to obtain the power generation under ash accumulation conditions. The difference between the theoretical clean power generation and the power generation under the ash accumulation state is calculated to obtain the real-time power loss value. The real-time power loss value is then integrated with a preset prediction time period to obtain the theoretical power generation loss value for each photovoltaic subarray.

4. The dynamic cleaning decision-making method for photovoltaic power plants based on digital twins and multi-objective optimization according to claim 1, characterized in that, The steps for constructing a multi-objective cleaning decision optimization model include: Based on the time-of-use electricity price data, the decision-making time period is determined. The on-grid electricity price coefficient; The water consumption cost and equipment wear and depreciation cost for a single cleaning operation are determined based on the status data of the cleaning equipment. Constructing the power generation revenue function and cleaning cost function And set the optimization objective as maximizing And minimize Power generation revenue function and cleaning cost function The function expressions are as follows: in, This indicates the total number of photovoltaic subarrays. Indicates the first The theoretical power generation loss value of each photovoltaic sub-array within the predicted time period. Indicates the first The power generation recovery rate of each photovoltaic subarray after cleaning Indicates the decision-making period The aforementioned on-grid electricity price coefficient, This represents the unit price of water resources per unit volume. This indicates the total water usage during the cleaning process. This indicates the wear and tear cost per unit time for cleaning equipment. This indicates the total working time of the cleaning equipment.

5. The photovoltaic power plant dynamic cleaning decision-making method based on digital twin and multi-objective optimization according to claim 4, characterized in that, The steps for constructing a multi-objective cleaning decision optimization model also include: Extract precipitation probability values ​​and wind speed prediction values ​​from the short-term weather forecast data; Set precipitation probability thresholds and wind speed safety thresholds; A first constraint condition is established, which is that when the precipitation probability value is greater than the precipitation probability threshold, the cleaning time window sequence is forcibly set to a non-working state. A second constraint is constructed, which is to stop generating the cleaning path planning instruction when the predicted wind speed value is greater than the wind speed safety threshold. The first constraint and the second constraint are added to the multi-objective cleaning decision optimization model.

6. The dynamic cleaning decision-making method for photovoltaic power plants based on digital twins and multi-objective optimization according to claim 1, characterized in that, The steps for obtaining dust monitoring data on the surface of the component include: Collect multispectral image data of the photovoltaic module surface; Texture features are extracted from the multispectral image data using a convolutional neural network model to obtain dust type classification results, wherein the dust type classification results include sand dust type, pollen type and oil stain type; Determine the corresponding dust adhesion coefficient based on the dust type classification results; The dust adhesion coefficient is used as part of the dust monitoring data on the component surface to correct the dust transmittance parameter in the digital twin model of the photovoltaic power station.

7. The dynamic cleaning decision-making method for photovoltaic power plants based on digital twins and multi-objective optimization according to claim 6, characterized in that, The steps for generating dynamic cleaning decision plans also include: Match the corresponding cleaning medium parameters according to the dust type classification result. When the dust type classification result is the oil stain type, increase the cleaning fluid concentration parameter. Adjust the brush head rotation speed and travel speed parameters of the cleaning device according to the dust adhesion coefficient; The cleaning medium parameters, the brush head rotation speed parameters, and the travel speed parameters are encapsulated into the dynamic cleaning decision scheme.

8. The dynamic cleaning decision-making method for photovoltaic power plants based on digital twins and multi-objective optimization according to claim 1, characterized in that, The steps for generating the cleaning path planning instructions include: Based on the theoretical power generation loss value, the multiple sub-array areas within the photovoltaic power station are sorted to obtain a power generation benefit priority sequence; Obtain the remaining battery power and current location coordinates from the status data of the cleaning device; When solving the cleaning path planning instructions, the ant colony algorithm is used to search for the optimal traversal path in the power generation benefit priority sequence, wherein the optimal traversal path is optimized to cover the high power generation benefit area and have the shortest total moving distance. The optimal traversal path is converted into the cleaning path planning instruction.

9. The dynamic cleaning decision-making method for photovoltaic power plants based on digital twins and multi-objective optimization according to claim 4, characterized in that, The steps for updating the decision weight parameters include: The difference between the actual power generation gain data and the actual cleaning cost data is calculated to obtain the actual net profit value; The expected net profit is calculated based on the power generation revenue and cleaning cost predicted by the multi-objective cleaning decision optimization model. Calculate the deviation rate between the actual net income value and the expected net income value; Construct a reinforcement learning reward function, and output a positive reward value when the deviation rate is less than a preset deviation threshold; The decision weight parameters in the multi-objective cleaning decision optimization model are updated using the positive reward value, wherein the decision weight parameters include power generation revenue weight and cleaning cost weight; The steps for solving the multi-objective cleaning decision optimization model using a multi-objective genetic algorithm include: Initialize a cleaning strategy population, wherein the cleaning strategy population includes several randomly generated combinations of cleaning time and path; According to the power generation revenue function and the cleaning cost function Calculate the non-dominated ranking rank and crowding distance of each individual in the population of the cleaning strategy; The cleaning strategy population is subjected to selection, crossover, and mutation operations to generate offspring populations; Merge the sweeping strategy population and the offspring population to select the Pareto optimal front solution set; Obtain the current decision weight parameters of the multi-objective cleaning decision optimization model, and denote them as the power generation revenue weight. and cleaning cost weighting ; Based on the power generation revenue weight and the cleaning cost weight Calculate the comprehensive decision score for each solution in the Pareto optimal frontier solution set. The formula for calculating the comprehensive decision score is as follows: The solution with the highest comprehensive decision score is selected as the dynamic cleaning decision scheme.

10. The system for dynamic cleaning decision-making of photovoltaic power plants based on digital twins and multi-objective optimization according to any one of claims 1-9, characterized in that, include: Digital twin modeling module: used to collect geographic information data, component layout topology data and electrical connection relationship data of photovoltaic power station, and to construct a digital twin model of photovoltaic power station based on the geographic information data, component layout topology data and electrical connection relationship data; Dust accumulation simulation and power generation loss assessment module: used to acquire real-time irradiance data and dust monitoring data on the surface of the components, input the real-time irradiance data and the dust monitoring data on the surface of the components into the digital twin model of the photovoltaic power station, simulate the real-time dust accumulation distribution state of the photovoltaic components, and calculate the theoretical power generation loss value based on the real-time dust accumulation distribution state of the photovoltaic components. Multi-objective cleaning optimization modeling module: used to acquire short-term weather forecast data, time-of-use electricity price data and cleaning equipment status data, and to construct a multi-objective cleaning decision optimization model based on the theoretical power generation loss value, the short-term weather forecast data, the time-of-use electricity price data and the cleaning equipment status data. The multi-objective cleaning decision optimization model is configured with decision weight parameters and takes maximizing power generation revenue and minimizing cleaning costs as optimization objectives. Dynamic cleaning decision generation module: used to solve the multi-objective cleaning decision optimization model using a multi-objective genetic algorithm, generate a dynamic cleaning decision scheme, wherein the dynamic cleaning decision scheme includes a cleaning time window sequence and a cleaning path planning instruction, and executes the cleaning operation according to the dynamic cleaning decision scheme; Model Adaptive Optimization Module: Used to collect actual power generation gain data and actual cleaning cost data after the cleaning operation is performed, calculate the decision reward value based on the actual power generation gain data and the actual cleaning cost data, and use a reinforcement learning algorithm to update the decision weight parameters of the multi-objective cleaning decision optimization model based on the decision reward value.

Images