Photovoltaic building self-adaptive cleaning system and method based on real-time carbon accounting

Abstract

The application discloses a photovoltaic building self-adaptive cleaning system and method based on real-time carbon accounting, and relates to the field of building integrated photovoltaic (BIPV) and low-carbon operation and maintenance technology.The application comprises a data sensing layer, a carbon accounting and decision-making layer and an intelligent execution layer.The data sensing layer collects multi-dimensional data such as power generation, environment and pollution at a high frequency;the carbon accounting and decision-making layer calculates real-time net carbon emission reduction benefits through a dynamic carbon accounting model, and outputs optimal cleaning decisions in combination with a deep reinforcement learning model;the intelligent execution layer executes cleaning operations according to the lowest carbon consumption path, and realizes continuous optimization of the model through data feedback.The application takes the maximization of net carbon benefits as a core target, breaks through the traditional fixed-cycle cleaning mode, realizes dynamic optimization of carbon benefits in the whole life cycle of the BIPV system, improves the carbon accounting precision, and significantly reduces operation and maintenance carbon emissions, and has wide popularization and application value.

Description

Technical Field

[0001] This invention belongs to the field of building-integrated photovoltaics (BIPV) and low-carbon operation and maintenance technology, and in particular relates to a photovoltaic building adaptive cleaning system and method based on real-time carbon accounting. Background Technology

[0002] Building-integrated photovoltaics (BIPV) is a key technology for promoting carbon neutrality in the building sector. Photovoltaic modules are directly integrated into the building's curtain wall or roof, enabling the integration of power generation and building functions, effectively improving building energy self-sufficiency.

[0003] However, the existing BIPV system's operation and maintenance technology has significant flaws, resulting in a large deviation between the actual emission reduction benefits and the theoretical values:

[0004] (1) Operation and maintenance carbon emissions are ignored: Traditional photovoltaic cleaning strategies are based on fixed cycles or simple power generation thresholds, without considering the carbon emissions generated by the cleaning behavior itself. Carbon emissions are generated in the cleaning process, such as water pump power consumption, cleaning agent production, and personnel transportation. In some low-pollution scenarios, "over-cleaning" will increase the overall carbon footprint of the system, which goes against the original intention of low-carbon construction.

[0005] (2) Lack of dynamic optimization in cleaning strategies: Existing technologies cannot comprehensively consider dynamic factors such as the spatiotemporal differences in dust accumulation rates, short-term weather forecasts, and real-time carbon emission intensity of regional power grids. For example, cleaning before rainfall can lead to the washing away of cleaning effects by rainwater, resulting in resource waste; cleaning during periods of high carbon intensity in the power grid can increase additional carbon emissions due to power consumption, making it difficult to achieve a globally optimal balance between "power generation loss costs" and "clean carbon costs".

[0006] (3) Static carbon accounting methods: Existing BIPV carbon accounting methods mostly adopt the static average value of life cycle assessment (LCA), such as the method recommended by the IEC 63216 standard. This method uses the annual average grid emission factor, which cannot reflect the interaction between real-time power generation emission reduction benefits and dynamic grid carbon intensity. The carbon footprint data update cycle is long, making it difficult to support refined and high-frequency carbon management decisions.

[0007] Therefore, there is an urgent need in this field for a system and method that can dynamically optimize the carbon benefits of BIPV systems throughout their entire lifecycle and intelligently drive operation and maintenance decisions, in order to solve the technical bottlenecks of traditional operation and maintenance technologies, such as extensive carbon accounting, simple decision-making mechanisms, and inefficient resource utilization. Summary of the Invention

[0008] The purpose of this invention is to provide a photovoltaic building adaptive cleaning system and method based on real-time carbon accounting. By using real-time and dynamic net carbon benefit calculation as the core criterion for cleaning decisions, it breaks through the traditional operation and maintenance mode based on time or efficiency thresholds, realizes the cognitive upgrade of BIPV system from power generation functional unit to carbon asset unit, and solves the problems of extensive carbon accounting, single decision-making mechanism and inefficient resource utilization in traditional operation and maintenance technology.

[0009] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0010] This invention relates to a photovoltaic building adaptive cleaning method based on real-time carbon accounting, comprising the following steps:

[0011] Step S1: Collect photovoltaic system operation data, environmental data, surface contamination data, and visual image data at a frequency of no less than 5 minutes per time through the data sensing layer; the operation data includes voltage, current, and power; the environmental data includes light intensity, temperature, humidity, wind speed, and weather forecast data for the next 24 hours; the contamination data includes photovoltaic surface transmittance and dust deposition density;

[0012] Step S2: Calculate the real-time net carbon reduction benefits of the system based on the dynamic carbon accounting model. The calculation logic includes the calculation of avoided carbon emissions, the separate calculation of process carbon emissions, and the coupled calculation of net carbon benefits.

[0013] Step S3: Input net carbon reduction benefits, pollution data, and meteorological forecast data into the depth. Network reinforcement learning model, the model is based on the state space and reward function Output cleaning decision instructions;

[0014] Step S4: If the decision instruction is to perform cleaning, the drive controller improves... The algorithm plans the cleaning path with the lowest carbon consumption and controls the cleaning execution device to complete the cleaning of fixed points or areas.

[0015] Step S5: Collect power generation improvement data and carbon cost data after cleaning, and feed the data back to the adaptive cleaning decision module for in-depth analysis. Online fine-tuning and optimization of the network model forms a closed-loop operation and maintenance system.

[0016] As a preferred technical solution, in step S1, the photovoltaic system operation data, environmental data, surface pollution data, and visual image data are calibrated and fused to output a dataset. The specific implementation process is as follows:

[0017] Step S11, Equipment Deployment: Deploy the power generation monitoring unit, environmental sensor group, surface contamination sensor and visual acquisition unit, and calibrate the parameters of the equipment at the same time;

[0018] Step S12, Data Acquisition: The current and voltage sensors collect the voltage across the components in real time. ), loop current ( ), based on real-time voltage ( ) and loop current ( ) Calculate instantaneous power ( ), and accumulate the total power generation within 5 minutes ( The data processing module processes the voltage within 5 minutes. ), loop current ( ) and instantaneous power ( Denoising (removing abnormal spikes) is performed, and the average value is taken as the running data for that period.

[0019] Step S13, Environmental Data Acquisition: The light intensity sensor collects the actual irradiance on the surface of the photovoltaic module (unit: The temperature and humidity sensor collects the surface temperature of the component (different from the ambient air temperature, error ±0.3℃) and relative humidity (error ±2%), while the wind speed sensor collects the horizontal wind speed (unit: It also records wind direction and calls a third-party meteorological platform (such as China Weather Network) via API to obtain 24-hour weather forecast data, including the probability of rainfall (%) and wind speed range (%). ), Irradiance prediction value ( );

[0020] Step S14, Surface Contamination Data Acquisition: The optical transmittance monitor emits a beam of light at a specific wavelength (550nm, the center wavelength of visible light), which penetrates the surface of the photovoltaic module and is detected by the receiver to calculate the current transmittance. ), and the initial transmittance of the component ( By comparing the transmittance attenuation rate (which is typically 95% at the factory calibration value) with the transmittance attenuation rate, we can obtain the transmittance attenuation rate. The current dust deposition density is calculated by mapping the transmittance attenuation rate to the dust deposition density. );

[0021] Step S15: Visual image data acquisition: A wide-angle high-definition camera captures images of the photovoltaic facade at a preset angle. The image format is JPG with a resolution of 1920×1080. The image data is transmitted to the data processing module for preprocessing (cropping and grayscale conversion) for subsequent identification of heavily polluted areas (such as bird droppings and stains).

[0022] Step S16, Data Transmission: Standardized data is transmitted to the carbon accounting and decision-making layer via industrial Ethernet or 5G; before transmission, the data is processed using… The criteria remove outlier data (such as out-of-range data caused by sensor malfunction or spike data caused by transient interference), supplement missing data using linear interpolation, and convert all data to a unified format. The format includes fields such as data type, collection timestamp, device ID, and measurement value.

[0023] As a preferred technical solution, in step S2, the calculation logic of the dynamic carbon accounting model includes:

[0024] Avoid carbon emissions : In the formula, for Photovoltaic power generation per period, in kWh (calculated from voltage and current data collected by the power generation monitoring unit). ;in, For component voltage, For component current, (sampling time interval) for Marginal emission factor of power grid in time period and region, unit (Reflects the carbon emission intensity of 1 kWh of newly generated electricity in the current power grid, obtained in real time from the power grid API, such as the peak hour carbon emission intensity of the eastern China power grid.) Valley time This factor is obtained from the power dispatch data platform via API; constraints: Furthermore, the update frequency is no less than once every 15 minutes to ensure real-time performance;

[0025] Process carbon emissions : In the formula, ; Water consumption for cleaning Carbon emissions from tap water production and transportation, based on data released by the local water company. The water pump consumes electricity for cleaning;

[0026] Net carbon emission reduction benefits : In the formula, This refers to the system's lifecycle.

[0027] As a preferred technical solution, in step S3, the depth Network reinforcement learning models are based on the state space and reward function The specific process for issuing cleaning decision instructions is as follows:

[0028] Step S31: Input net carbon emission reduction benefits Photovoltaic surface transmittance Probability of rainfall in the next 24 hours Wind speed in the next 24 hours Current power grid carbon intensity Rainwater tank water level Preprocessing operations are performed.

[0029] Step S32: Encode the preprocessed data into a 5-dimensional state vector that can be recognized by the DQN model. ( (At the current moment):

[0030] ;

[0031] Step S33: The DQN model adopts a "convolutional layer + fully connected layer" structure, with a 5-dimensional state vector as input. The output is a 3D motion value. Values; the network structure is as follows:

[0032] Input layer: 5 neurons (corresponding to the dimension of the state vector);

[0033] Hidden layer 1: 64 neurons, activation function ReLU;

[0034] Hidden layer 2: 32 neurons, activation function ReLU;

[0035] Output layer: 3 neurons (corresponding to the 3 actions in action space A), outputting the value of each action. value ( (for the i-th action).

[0036] Step S34: The DQN model is iteratively updated using the Bellman equation. The value, which maximizes long-term rewards, is calculated using the following formula:

[0037] ;

[0038] In the formula, Current state Next action The value of the action, As an immediate reward for the current action, Discount factor (values) (The value of 0.95 in this invention represents the degree of importance attached to future rewards). For the next state The maximum Q value for all actions;

[0039] Step S35: Set the exploration rate If random number If the random number is selected, then the action is randomly selected; if the random number is selected, then the action is randomly selected. Then choose The biggest movement;

[0040] Step S36: Based on the selected action, output a clear decision instruction, the specific decision instruction is as follows:

[0041] Action 0 (Not Clean): Output "Maintain the current state, re-collect data and recalculate in 5 minutes";

[0042] Action 1 (Immediate Cleaning): Output "Start cleaning device and perform cleaning along the lowest carbon consumption path";

[0043] Action 2 (Delayed Cleaning): Output "Delay the assessment to the next time period (e.g., 1 hour later), and continuously monitor changes in contamination during the period."

[0044] As a preferred technical solution, the reward function The specific formula is:

[0045] ;

[0046] In the formula, This represents the expected increase in net carbon emissions after the cleanup action. , and These are the weighting coefficients.

[0047] As a preferred technical solution, in step S4, by improving... The specific process of the algorithm to plan the clean path with the lowest carbon consumption is as follows:

[0048] Step S41: Construct a photovoltaic facade grid coordinate system: Divide the photovoltaic building facade into two-dimensional grid nodes according to the working units of the cleaning device, and define the coordinates of each node as follows: ;in, The coordinates are horizontal (unit: meters), and the range of values ​​is the total horizontal length of the photovoltaic facade. The coordinates are in the vertical direction (unit: meters), and the values ​​range from the total vertical height of the photovoltaic facade; each grid node represents the smallest unit where the cleaning device can stay / operate (e.g., Heavily polluted areas (such as bird droppings and stains) are marked as core operation nodes, and moderately polluted areas are marked as regular operation nodes.

[0049] Define adjacent nodes (up, down, left, right) as reachable nodes, and non-adjacent nodes as nodes that need to be jumped through intermediate nodes. At the same time, mark structural obstacle nodes (such as window frames and columns) as unreachable nodes.

[0050] Step S42: Calculate carbon consumption weights between nodes: Assign carbon consumption weights to the paths between every two reachable nodes. The weight calculation formula is as follows:

[0051] ;

[0052] In the formula, The carbon consumption weight is used for node-to-node carbon consumption. For cleaning devices from nodes To the node The mobile power consumption is calculated using the following formula: ; for The marginal emission factor of the power grid during the time period Let i be the cleaning water consumption of the area covered by the path from node i to j. Carbon emission factors for tap water / rainwater The rated power of the cleaning device, For nodes To the node The straight-line distance The moving speed of the cleaning device; The calculation formula is: ; This represents the photovoltaic area corresponding to a single node. Water consumption per unit area for cleaning;

[0053] Step S43, Improvement Algorithm initialization: Set the start node and end node, and initialize the parameters; set the start node to the initial docking position of the cleaning device (e.g., the bottom left side of the photovoltaic facade), denoted as . The endpoint node is defined as the return position (e.g., the initial docking position) after all core task nodes and regular task nodes have been traversed, denoted as . ;

[0054] Step S44, Improvement The algorithm iteratively solves for the optimal path: iterates through all reachable adjacent nodes, calculates the carbon consumption of the path, selects the path with the minimum carbon consumption, and generates the optimal path;

[0055] Step S45, Cleaning Execution: The drive controller converts the optimal path into motion commands for the cleaning device (such as the number of stepper motor pulses for the track robot and the flight coordinates of the drone). Core operation nodes perform deep cleaning (increasing water consumption / cleaning time), and routine operation nodes perform routine cleaning. During the operation, the device's energy consumption and water consumption data are collected in real time and fed back to the carbon accounting module to correct subsequent weight calculations.

[0056] As a preferred technical solution, in step S44, the improvement The specific process of the algorithm iteratively finding the optimal path is as follows:

[0057] Step S441: Select the node with the lowest carbon consumption among the unvisited nodes. ;

[0058] Step S442: Transfer the node Add to visited collection ,from Remove from;

[0059] Step S443: Traverse All reachable adjacent nodes ;

[0060] Step S444: Calculate from start to Then Carbon consumption;

[0061] Step S445: If the new carbon consumption is less than the current ,renew and predecessor nodes;

[0062] Step S446: Backtrack the predecessor node to generate the optimal path and obtain the path with the lowest carbon consumption from start to end;

[0063] To address the issue of "turning back" during traversal, a greedy strategy is employed for optimization:

[0064] Prioritize traversing core operational nodes within the same vertical / horizontal direction to reduce cross-regional movement; if adjacent nodes have different carbon consumption weights... Prioritize shorter routes (balancing carbon consumption and efficiency).

[0065] As a preferred technical solution, in step S5, the depth The specific process for online fine-tuning and optimization of the network model is as follows:

[0066] Step S51, Cleaning Effect and Carbon Cost Data Acquisition: High-frequency acquisition of photovoltaic operation data and resource consumption data within 24 hours after cleaning; Inputs: Light transmittance after cleaning, hourly power generation, actual water consumption / power consumption, cleaning duration; Output: Raw feedback dataset ;

[0067] Step S52, Data Preprocessing and Labeling: Calculate the actual carbon benefit improvement and carbon cost deviation, and remove outliers; Input: Expected values ​​calculated before cleaning ( Output: Standardized feedback dataset ;

[0068] Step S53, Online Fine-tuning of the DQN Model: Update model weights based on labeled data and optimize the reward function weight coefficients; Input: Original model parameters Output: New model parameters Optimized weight coefficients ;

[0069] Step S54, Fine-tuning and solidifying parameters: Verify the decision-making effect of the new parameters; if they meet the threshold, solidify the parameters. Input: , Validation set data; Output: Fixed model parameters ;

[0070] Step S55, Closed-loop verification: Compare the carbon benefits before and after fine-tuning to confirm the optimization effect; Input: Carbon benefit data for 10 decisions before and after fine-tuning; Output: Closed-loop optimization effect report.

[0071] As a preferred technical solution, in step S53, the loss function calculation formula for the DQN model is as follows:

[0072] ;

[0073] In the formula, The target value is calculated using the following formula: ; For the current model, the state ,action Valuation of value; For the target network parameters, Discount factor;

[0074] The formula for optimizing the weighting coefficients is: In the formula, This is the first weighting coefficient after optimization. This is the learning rate.

[0075] This invention is a photovoltaic building adaptive cleaning system based on real-time carbon accounting, comprising a data sensing layer, a carbon accounting and decision-making layer, and an intelligent execution layer. The three layers sequentially complete data acquisition, carbon accounting and decision generation, and cleaning execution and feedback optimization.

[0076] The data sensing layer is deployed on the photovoltaic building facade and includes a power generation monitoring unit, an environmental sensor group, a surface contamination sensor, and a visual acquisition unit. The power generation monitoring unit includes a current and voltage sensor and a power meter. The environmental sensor group includes a light intensity sensor, a temperature sensor, a humidity sensor, and a wind speed sensor. The surface contamination sensor is an optical transmittance monitor with a measurement accuracy of ±0.5%. The visual acquisition unit is at least one wide-angle high-definition camera with low-light imaging capability.

[0077] The carbon accounting and decision-making layer includes a data processing module, a dynamic carbon accounting module, and an adaptive clean decision-making module. The dynamic carbon accounting module has an updatable life cycle carbon emission factor library built-in and connects to the regional power grid real-time / predictive marginal emission factor data platform through an API interface to calculate the real-time net carbon emission reduction benefits of the photovoltaic system. The adaptive clean decision-making module integrates a deep Q-network reinforcement learning model and outputs clean decision instructions with the long-term goal of maximizing net carbon emission reduction benefits.

[0078] The intelligent execution layer includes a cleaning execution device and a drive controller; the cleaning execution device is a low-water-consumption or water-free device, selected from one or more of the following: a track-type cleaning robot, a drone cleaning system, and a local spot spraying device; the drive controller receives instructions from the decision layer and controls the cleaning execution device to perform fixed-point or area cleaning according to the planned path.

[0079] The present invention has the following beneficial effects:

[0080] (1) This invention realizes carbon signal-driven operation and maintenance closed loop, and takes real-time and dynamic net carbon benefit calculation as the core judgment basis for clean decision-making. It breaks through the traditional operation and maintenance mode based on time or efficiency threshold, realizes the cognitive upgrade of BIPV system from power generation functional unit to carbon asset unit, and improves carbon accounting accuracy in typical application scenarios.

[0081] (2) By constructing an updatable factor library that includes manufacturing, transportation, installation, operation and maintenance and recycling, and linking it with real-time power generation carbon emission reduction, this invention realizes high-frequency (hourly) and high-precision (error <3%) calculation of carbon flow in BIPV system, providing a real-time data foundation for carbon trading and green building certification, and realizing refined carbon management throughout the entire life cycle.

[0082] (3) This invention designs a reinforcement learning reward function with carbon benefits as its core, intelligently weighs multiple factors such as power generation gain, water resource consumption, and power consumption, and automatically seeks the best in a complex and changing environment, ensuring power generation efficiency while reducing clean water consumption and reducing carbon emissions from operation and maintenance.

[0083] (4) The present invention adopts a modular design, the carbon accounting model can be connected with the power grid carbon data platform in different regions, and the decision algorithm can be adapted to different building structures (curtain walls, roofs) and cleaning devices (robots, drones, sprinkler systems), which has broad application value and improves system compatibility and scalability.

[0084] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0085] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0086] Figure 1 This is a flowchart of a photovoltaic building adaptive cleaning method based on real-time carbon accounting according to the present invention.

[0087] Figure 2 This is a schematic diagram of a photovoltaic building adaptive cleaning system based on real-time carbon accounting according to the present invention.

[0088] Figure 3 For depth Network reinforcement learning models are based on the state space and reward function It outputs cleaning decision instructions. Detailed Implementation

[0089] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0090] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0091] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-3 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0092] Example 1

[0093] Please see Figure 1 As shown, this invention is a photovoltaic building adaptive cleaning method based on real-time carbon accounting, comprising the following steps:

[0094] Step S1: Collect photovoltaic system operation data, environmental data, surface contamination data, and visual image data at a frequency of no less than 5 minutes / time through the data sensing layer; operation data includes voltage, current, and power; environmental data includes light intensity, temperature, humidity, wind speed, and weather forecast data for the next 24 hours; contamination data includes photovoltaic surface transmittance and dust deposition density;

[0095] Step S2: Calculate the real-time net carbon reduction benefits of the system based on the dynamic carbon accounting model. The calculation logic includes the calculation of avoided carbon emissions, the separate calculation of process carbon emissions, and the coupled calculation of net carbon benefits.

[0096] Step S3: Input net carbon reduction benefits, pollution data, and meteorological forecast data into the depth. Network reinforcement learning model, the model is based on the state space and reward function Output cleaning decision instructions;

[0097] Step S4: If the decision instruction is to perform cleaning, the drive controller improves... The algorithm plans the cleaning path with the lowest carbon consumption and controls the cleaning execution device to complete the cleaning of fixed points or areas.

[0098] Step S5: Collect power generation improvement data and carbon cost data after cleaning, and feed the data back to the adaptive cleaning decision module for in-depth analysis. Online fine-tuning and optimization of the network model forms a closed-loop operation and maintenance system.

[0099] In step S1, the photovoltaic system operation data, environmental data, surface contamination data, and visual image data are calibrated and fused to output a dataset. The specific implementation process is as follows:

[0100] Step S11, Equipment Deployment: Deploy the power generation monitoring unit, environmental sensor group, surface contamination sensor and visual acquisition unit, and calibrate the parameters of the equipment at the same time;

[0101] Specifically, the equipment is deployed in a grid-like manner as follows:

[0102] Power generation monitoring unit: Press " "The photovoltaic modules are evenly distributed on the building facade to ensure that the area covered by each monitoring unit is unobstructed. The current and voltage sensors are precisely connected to the positive and negative terminals of the modules, and the power meter is linked to the building's power distribution system."

[0103] Environmental sensor array: The array is deployed in a "triangular distribution" (one set each at the top, middle and bottom of the building), with the sensor probes facing the sun-receiving surface of the photovoltaic modules to avoid shading themselves; the wind speed sensor is installed 1m above the top of the photovoltaic facade to reduce turbulence interference.

[0104] Surface contamination sensors: Deployed one-to-one with the power generation monitoring unit. The transmitter and receiver of the optical transmittance monitor are installed on the two sides of the photovoltaic module respectively to ensure that the light path is perpendicular to the module surface and there is no dust accumulation interference.

[0105] Visual acquisition unit: Deployed at the highest point of the building (such as the rooftop), with the camera lens focal length adjusted to cover the entire photovoltaic facade, and the shooting angle perpendicular to the facade. Avoid shooting against the light; turn on low light mode and automatic exposure compensation to ensure image clarity on cloudy days / nights.

[0106] The calibration method for the equipment parameters is as follows:

[0107] Power generation monitoring unit: By inputting known voltage and current signals (such as DC36V, 10A) through a standard power simulator, the sensor measurement error is calibrated to ensure voltage measurement accuracy of ±0.5%, current measurement accuracy of ±0.3%, and power calculation error of <1%.

[0108] Surface contamination sensor: Calibrated using standard transmittance calibration plates (90%, 80%, 70% transmittance) to correct measurement deviations and ensure final transmittance measurement accuracy of ±0.5%.

[0109] Visual acquisition unit: White balance and contrast calibration are performed by shooting a standard grayscale card, and a fixed shooting resolution (1920×1080) and frame rate (1 frame / 5 minutes) are set to avoid data redundancy.

[0110] Step S12, Data Acquisition: The current and voltage sensors collect the voltage across the components in real time. ), loop current ( ), based on real-time voltage ( ) and loop current ( ) Calculate instantaneous power ( ), and accumulate the total power generation within 5 minutes ( The data processing module processes the voltage within 5 minutes. ), loop current ( ) and instantaneous power ( Denoising (removing abnormal spikes) is performed, and the average value is taken as the running data for that period.

[0111] Specifically, the formula for calculating the total power generation within 5 minutes is as follows:

[0112] ;

[0113] In the formula, The total power generation (kWh) during time period t (5 minutes) For the first Instantaneous power per second (W), 300 represents the number of seconds in a 5-minute period, and 3600 is the coefficient for converting power (W) to energy (kWh);

[0114] Step S13, Environmental Data Acquisition: The light intensity sensor collects the actual irradiance on the surface of the photovoltaic module (unit: The temperature and humidity sensor collects the surface temperature of the component (different from the ambient air temperature, error ±0.3℃) and relative humidity (error ±2%), while the wind speed sensor collects the horizontal wind speed (unit: It also records wind direction and calls a third-party meteorological platform (such as China Weather Network) via API to obtain 24-hour weather forecast data, including the probability of rainfall (%) and wind speed range (%). ), Irradiance prediction value ( );

[0115] Step S14, Surface Contamination Data Acquisition: The optical transmittance monitor emits a beam of light at a specific wavelength (550nm, the center wavelength of visible light), which penetrates the surface of the photovoltaic module and is detected by the receiver to calculate the current transmittance. ), and the initial transmittance of the component ( By comparing the transmittance attenuation rate (which is typically 95% at the factory calibration value) with the transmittance attenuation rate, we can obtain the transmittance attenuation rate. The current dust deposition density is calculated by mapping the transmittance attenuation rate to the dust deposition density. );

[0116] Specifically, the formula for calculating the transmittance attenuation rate is: In the formula, Transmittance attenuation rate ( ), This refers to the initial light transmittance of the photovoltaic module (factory rated value, default 95%). The current measured transmittance (%).

[0117] Dust deposition density ( The calculation formula (based on the experimentally calibrated mapping model) is as follows:

[0118] ;

[0119] In the formula, Dust deposition density (g / ), Calibration coefficients (obtained experimentally, with a range of values) ), The correction factor (obtained experimentally, with a range of values) is... ).

[0120] Note: and The calculation must be calibrated experimentally based on the surface material of the photovoltaic module (such as glass or thin film) and the type of local dust (such as sand or dust) to ensure accuracy in determining the calculation error. .

[0121] Step S15: Visual image data acquisition: A wide-angle high-definition camera captures images of the photovoltaic facade at a preset angle. The image format is JPG with a resolution of 1920×1080. The image data is transmitted to the data processing module for preprocessing (cropping and grayscale conversion) for subsequent identification of heavily polluted areas (such as bird droppings and stains).

[0122] Specifically, for high-frequency data such as current and voltage, a moving average method is used to reduce random interference: ;

[0123] In the formula, for The average value after noise reduction at any given time. The sliding window size (value 10, meaning the average of the raw data over 10 consecutive seconds). for The raw data at any given moment.

[0124] Step S16, Data Transmission: Standardized data is transmitted to the carbon accounting and decision-making layer via industrial Ethernet or 5G, with a transmission delay of <1 second and a data packet loss rate of <0.1%; before transmission, the following steps are taken: The criteria remove outlier data (such as out-of-range data caused by sensor malfunction or spike data caused by transient interference), supplement missing data using linear interpolation, and convert all data to a unified format. The format includes fields such as data type, collection timestamp, device ID, and measurement value.

[0125] In specific implementation, after deploying the data sensing layer of this invention on the BIPV curtain wall of a 20-story high-rise building (area 1200㎡, component power 200kW), the data acquisition process of step S1 is executed:

[0126] The carbon accounting and decision-making layer sent a data collection trigger signal at 10:00:00 on December 1, 2025, and all sensing devices started data collection simultaneously.

[0127] Operational data acquisition: Current and voltage sensors acquire data every 1 second, for a total of 300 sets of data within 5 minutes, including voltage data range. Current data range ;

[0128] After noise reduction using the moving average method, the average voltage is 35.8V and the average current is 9.2A.

[0129] Calculate the average instantaneous power: ;

[0130] Calculate the amount of electricity generated in 5 minutes: .

[0131] Environmental data collection is as follows: Values ​​collected by the light intensity sensor: Component surface temperature: 29.3℃, relative humidity: 62%; wind speed: 2.3m / s, wind direction: southwest;

[0132] The weather forecast data is as follows: 25% probability of rainfall in the next 24 hours, average wind speed... average irradiance .

[0133] Surface contamination data was collected as follows: initial light transmittance of the component. The current measured transmittance T = 78%; calculate the transmittance attenuation rate: ;

[0134] The local dust type is sandy soil, experimental calibration , ;

[0135] Calculate dust deposition density: .

[0136] The visual image data was collected as follows: Images of the photovoltaic curtain wall were captured by a rooftop camera. After preprocessing, two heavily polluted areas (bird droppings, each approximately 0.8 square meters in size) were identified in the center of the curtain wall. Data output: All data, standardized to JSON format, is transmitted to the carbon accounting and decision-making level with a transmission delay of 0.8 seconds and no data loss. The final output is... Time period ( ) dataset.

[0137] Through the above process, step S1 achieves multi-dimensional, high-precision, and timed data collection, providing reliable data support for subsequent dynamic carbon accounting and clean energy decision-making.

[0138] In step S2, the calculation logic of the dynamic carbon accounting model includes:

[0139] Avoid carbon emissions : In the formula, for Photovoltaic power generation per period, in kWh (calculated from voltage and current data collected by the power generation monitoring unit). ;in, For component voltage, For component current, (sampling time interval) for Marginal emission factor of power grid in time period and region, unit (Reflects the carbon emission intensity of 1 kWh of newly generated electricity in the current power grid, obtained in real time from the power grid API, such as the peak hour carbon emission intensity of the eastern China power grid.) Valley time This factor is obtained from the power dispatch data platform via API; constraints: Furthermore, the update frequency is no less than once every 15 minutes to ensure real-time performance;

[0140] Process carbon emissions : In the formula, ; Water consumption for cleaning Carbon emissions from tap water production and transportation, based on data released by the local water company. The water pump consumes electricity for cleaning;

[0141] Net carbon emission reduction benefits : In the formula, This refers to the system's lifecycle.

[0142] Specifically, the core objective of step S2 is to calculate the carbon emission of a building-integrated photovoltaic (BIPV) system based on real-time data from the data perception layer and static carbon emission data throughout its entire lifecycle (without allocation), through the logic of "avoiding carbon emission calculation, calculating process carbon emissions, and coupling net carbon benefits." Net carbon emission reduction benefits over time period This provides a quantitative basis for cleanliness decisions.

[0143] Data preprocessing and aggregation are as follows:

[0144] 1. Receive real-time data collected by the data sensing layer: photovoltaic power generation during time period t. Water consumption for planned / actual cleaning Power consumption of cleaning water pump ;

[0145] 2. Accessing external dynamic carbon data: Obtained via API Marginal emission factor of power grid in time period area Carbon emission factors in tap water ;

[0146] Read static basic data: retrieve total carbon emissions during the photovoltaic module manufacturing stage from the system database. Total carbon emissions during the transportation phase (One-time data entry, no need to allocate costs);

[0147] Unified data units: All data units are standardized (e.g., electricity generation). Carbon emissions: Water consumption: The time granularity is unified as follows: Time period (e.g., 1 hour / 5 minutes).

[0148] In practice, data is collected in real time. Photovoltaic power generation during the period It is 120kWh; the cleaning operation and maintenance parameters are: water consumption Water pump power If it runs for 0.5 hours, then ;

[0149] The distribution calculation process is as follows:

[0150] Calculate the carbon emissions to avoid : ;

[0151] Calculate carbon emissions from clean operation and maintenance : ;

[0152] Carbon emissions during calculation : ;

[0153] Calculate the net carbon emission reduction benefits : ;

[0154] The calculation results show that This indicates that the emission reduction from power generation during period t has not yet covered the carbon cost throughout the system's entire lifecycle. The decision-making model needs further analysis based on this result: If clean energy is implemented, the power generation is expected to increase to 150 kWh in the next period (11:00-12:00), corresponding to... ;

[0155] (The negative impact of net carbon benefits is narrowing, meaning the carbon cost recovery rate is accelerating.)

[0156] Therefore, the model's decision to "clean immediately" is based on the core logic that cleaning, although it increases the workload slightly... However, it can improve subsequent periods. Accelerate coverage To maximize long-term carbon benefits by minimizing static carbon costs.

[0157] Please see Figure 3 As shown, in step S3, the depth Network reinforcement learning models are based on the state space and reward function The specific process for issuing cleaning decision instructions is as follows:

[0158] Step S31: Input net carbon emission reduction benefits Photovoltaic surface transmittance Probability of rainfall in the next 24 hours Wind speed in the next 24 hours Current power grid carbon intensity Rainwater tank water level Preprocessing operations are performed.

[0159] Step S32: Encode the preprocessed data into a 5-dimensional state vector that can be recognized by the DQN model. ( (At the current moment):

[0160] ;

[0161] Step S33: The DQN model adopts a "convolutional layer + fully connected layer" structure, with a 5-dimensional state vector as input. The output is a 3D motion value. Values; the network structure is as follows:

[0162] Input layer: 5 neurons (corresponding to the dimension of the state vector);

[0163] Hidden layer 1: 64 neurons, activation function ReLU;

[0164] Hidden layer 2: 32 neurons, activation function ReLU;

[0165] Output layer: 3 neurons (corresponding to the 3 actions in action space A), outputting the value of each action. value ( (For the i-th action).

[0166] Step S34: The DQN model is iteratively updated using the Bellman equation. The value, which maximizes long-term rewards, is calculated using the following formula:

[0167] ;

[0168] In the formula, Current state Next action The value of the action, As an immediate reward for the current action, Discount factor (values) (The value of 0.95 in this invention represents the degree of importance attached to future rewards). For the next state The maximum Q value for all actions;

[0169] Step S35: Set the exploration rate If random number If the random number is selected, then the action is randomly selected; if the random number is selected, then the action is randomly selected. Then choose The biggest movement;

[0170] Step S36: Based on the selected action, output a clear decision instruction, the specific decision instruction is as follows:

[0171] Action 0 (Not Clean): Output "Maintain the current state, re-collect data and recalculate in 5 minutes";

[0172] Action 1 (Immediate Cleaning): Output "Start cleaning device and perform cleaning along the lowest carbon consumption path";

[0173] Action 2 (Delayed Cleaning): Output "Delay the assessment to the next time period (e.g., 1 hour later), and continuously monitor changes in contamination during the period";

[0174] In step S3, depth Network reinforcement learning models are based on the state space and reward function The specific process for issuing cleaning decision instructions is as follows:

[0175] Step S31: Input net carbon emission reduction benefits Photovoltaic surface transmittance Probability of rainfall in the next 24 hours Wind speed in the next 24 hours Current power grid carbon intensity Rainwater tank water level Preprocessing operations are performed.

[0176] Step S32: Encode the preprocessed data into a 5-dimensional state vector that can be recognized by the DQN model. ( (At the current moment):

[0177] ;

[0178] Step S33: The DQN model adopts a "convolutional layer + fully connected layer" structure, with a 5-dimensional state vector as input. The output is a 3D motion value. Values; the network structure is as follows:

[0179] Input layer: 5 neurons (corresponding to the dimension of the state vector);

[0180] Hidden layer 1: 64 neurons, activation function ReLU;

[0181] Hidden layer 2: 32 neurons, activation function ReLU;

[0182] Output layer: 3 neurons (corresponding to the 3 actions in action space A), outputting the value of each action. value ( (for the i-th action).

[0183] Step S34: The DQN model is iteratively updated using the Bellman equation. The value, which maximizes long-term rewards, is calculated using the following formula:

[0184] ;

[0185] In the formula, Current state Next action The value of the action, As an immediate reward for the current action, Discount factor (values) (The value of 0.95 in this invention represents the degree of importance attached to future rewards). For the next state The maximum Q value for all actions;

[0186] Step S35: Set the exploration rate If random number If the random number is selected, then the action is randomly selected; if the random number is selected, then the action is randomly selected. Then choose The biggest movement;

[0187] Step S36: Based on the selected action, output a clear decision instruction, the specific decision instruction is as follows:

[0188] Action 0 (Not Clean): Output "Maintain the current state, re-collect data and recalculate in 5 minutes";

[0189] Action 1 (Immediate Cleaning): Output "Start cleaning device and perform cleaning along the lowest carbon consumption path";

[0190] Action 2 (Delayed Cleaning): Output "Delay the assessment to the next time period (e.g., 1 hour later), and continuously monitor changes in contamination during the period."

[0191] reward function The specific formula is:

[0192] ;

[0193] In the formula, This represents the expected increase in net carbon emissions after the cleanup action. , and Weighting coefficients (satisfying) In this embodiment, the initial value is set to (Carbon efficiency first) (Carbon emission constraints) (Water resource constraints) will be dynamically adjusted according to training.

[0194] In practice, taking a 200kW BIPV curtain wall of a commercial building as an example, the following data was collected at time t (10:00):

[0195] Carbon Dioxide Benefits Before Cleaning Light transmittance (Standardized value: 0.5); Probability of rainfall in the next 24 hours (0.3 after standardization); wind speed (0.5 after standardization); Grid carbon intensity (0.6 after standardization); Water level in the water tank (Standardized value: 0.6); Model parameters: , , , , .

[0196] The specific process is as follows:

[0197] All data points are free of outliers. Standardized state vector: ;

[0198] The DQM model outputs the Q-values ​​for the three actions:

[0199] Action 0 (Not Clean): Action 1 (Immediate Cleaning): Action 2 (Delayed Cleaning): .

[0200] random numbers Select action 1 (clean immediately) with the highest Q value;

[0201] Predicted net carbon efficiency gains after cleaning Clean carbon emissions Water consumption for cleaning The reward function is calculated as follows: Verification showed that the Q-value (8.7) of action 1 and the reward value (8.9632) were highly matched, indicating that the decision was reasonable.

[0202] The model outputs an immediate cleaning command, which drives the controller to start the cleaning robot to perform the cleaning operation.

[0203] After cleaning, the light transmittance was restored to 91%. Power generation increased to Net carbon benefit value Actual reward value Model prediction error The validity of the decision has been verified.

[0204] In step S4, by improving The specific process of the algorithm to plan the clean path with the lowest carbon consumption is as follows:

[0205] Step S41: Construct a photovoltaic facade grid coordinate system: Divide the photovoltaic building facade into two-dimensional grid nodes according to the working units of the cleaning device, and define the coordinates of each node as follows: ;in, The coordinates are horizontal (unit: meters), and the range of values ​​is the total horizontal length of the photovoltaic facade. The coordinates are in the vertical direction (unit: meters), and the values ​​range from the total vertical height of the photovoltaic facade; each grid node represents the smallest unit where the cleaning device can stay / operate (e.g., Heavily polluted areas (such as bird droppings and stains) are marked as core operation nodes, and moderately polluted areas are marked as regular operation nodes.

[0206] Define adjacent nodes (up, down, left, right) as reachable nodes, and non-adjacent nodes as nodes that need to be jumped through intermediate nodes. At the same time, mark structural obstacle nodes (such as window frames and columns) as unreachable nodes.

[0207] Step S42: Calculate carbon consumption weights between nodes: Assign carbon consumption weights to the paths between every two reachable nodes. The weight calculation formula is as follows:

[0208] ;

[0209] In the formula, The carbon consumption weight is used for node-to-node carbon consumption. For cleaning devices from nodes To the node The mobile power consumption is calculated using the following formula: ; for The marginal emission factor of the power grid during the time period Let i be the cleaning water consumption of the area covered by the path from node i to j. Carbon emission factors for tap water / rainwater The rated power of the cleaning device, For nodes To the node The straight-line distance The moving speed of the cleaning device; The calculation formula is: ; This represents the photovoltaic area corresponding to a single node. Water consumption per unit area for cleaning;

[0210] Step S43, Improvement Algorithm initialization: Set the start node and end node, and initialize the parameters; set the start node to the initial docking position of the cleaning device (e.g., the bottom left side of the photovoltaic facade), denoted as . The endpoint node is defined as the return position (e.g., the initial docking position) after all core task nodes and regular task nodes have been traversed, denoted as . ;

[0211] Step S44, Improvement The algorithm iteratively solves for the optimal path: iterates through all reachable adjacent nodes, calculates the carbon consumption of the path, selects the path with the minimum carbon consumption, and generates the optimal path;

[0212] Step S45, Cleaning Execution: The drive controller converts the optimal path into motion commands for the cleaning device (such as the number of stepper motor pulses for the track robot and the flight coordinates of the drone). Core operation nodes perform deep cleaning (increasing water consumption / cleaning time), and routine operation nodes perform routine cleaning. During the operation, the device's energy consumption and water consumption data are collected in real time and fed back to the carbon accounting module to correct subsequent weight calculations.

[0213] In step S44, improvements are made. The specific process of the algorithm iteratively finding the optimal path is as follows:

[0214] Step S441: Select the node with the lowest carbon consumption among the unvisited nodes. ;

[0215] Step S442: Transfer the node Add to visited collection ,from Remove from;

[0216] Step S443: Traverse All reachable adjacent nodes ;

[0217] Step S444: Calculate from start to Then Carbon consumption;

[0218] Step S445: If the new carbon consumption is less than the current ,renew and predecessor nodes;

[0219] Step S446: Backtrack the predecessor node to generate the optimal path and obtain the path with the lowest carbon consumption from start to end;

[0220] To address the issue of "turning back" during traversal, a greedy strategy is employed for optimization:

[0221] Prioritize traversing core operational nodes within the same vertical / horizontal direction to reduce cross-regional movement; if adjacent nodes have different carbon consumption weights... Prioritize shorter routes (balancing carbon consumption and efficiency).

[0222] In practical implementation, the carbon consumption weighting between nodes is calculated (using...). (For example); distance Time taken to move: Mobile power consumption: Mobile carbon consumption: This node is a non-operating node, and its water consumption is... Water-related carbon consumption = 0; weighting .

[0223] Core task node weight calculation (taking N(20,15) as an example): Node area Water consumption Water-related carbon consumption: Assuming from The carbon consumption of mobile devices is Total weight .

[0224] The algorithm iteratively solves for the optimal path as follows: Initialization The remaining nodes are First iteration: Select Update its adjacent nodes , of The value is 0.000028; subsequent iterations: prioritize traversing the node with the lowest carbon consumption, gradually covering all core task nodes and regular task nodes; finally, the optimal path is generated: The total carbon consumption of this route is Compared to the traditional shortest distance path (total carbon consumption) )reduce .

[0225] In step S5, depth The specific process for online fine-tuning and optimization of the network model is as follows:

[0226] Step S51, Cleaning Effect and Carbon Cost Data Acquisition: High-frequency acquisition of photovoltaic operation data and resource consumption data within 24 hours after cleaning; Inputs: Light transmittance after cleaning, hourly power generation, actual water consumption / power consumption, cleaning duration; Output: Raw feedback dataset ;

[0227] Step S52, Data Preprocessing and Labeling: Calculate the actual carbon benefit improvement and carbon cost deviation, and remove outliers; Input: Expected values ​​calculated before cleaning ( Output: Standardized feedback dataset ;

[0228] Specifically, the actual net carbon benefit improvement: In the formula, After cleaning The actual net carbon emission reduction benefits over a given period of time Net carbon emission reduction benefits under the same operating conditions before cleaning (historical or projected values);

[0229] Carbon cost bias: In the formula, To clean up carbon emissions from actual production, Clean carbon emissions are estimated during the decision-making phase;

[0230] Actual reward value: ; The actual reward, based on real data, serves as a label for model fine-tuning.

[0231] Step S53, Online Fine-tuning of the DQN Model: Update model weights based on labeled data and optimize the reward function weight coefficients; Input: Original model parameters Output: New model parameters Optimized weight coefficients ;

[0232] Step S54, Fine-tuning and solidifying parameters: Verify the decision-making effect of the new parameters; if they meet the threshold, solidify the parameters. Input: , Validation set data; Output: Fixed model parameters ;

[0233] Specifically, , In the formula, This represents the number of times the fine-tuned decision aligns with the "optimal carbon benefit decision." For carbon efficiency deviation rate, it must be ≤5% before parameters are solidified.

[0234] Step S55, Closed-loop verification: Compare the carbon benefits before and after fine-tuning to confirm the optimization effect; Input: Carbon benefit data for 10 decisions before and after fine-tuning; Output: Closed-loop optimization effect report.

[0235] In step S53, the loss function of the DQN model is calculated as follows:

[0236] ;

[0237] In the formula, The target value is calculated using the following formula: ; For the current model, the state ,action Valuation of value; For the target network parameters, Discount factor;

[0238] The formula for optimizing the weighting coefficients is: In the formula, This is the first weighting coefficient after optimization. The learning rate;

[0239] If the actual carbon benefit improvement is higher than expected, increase (Carbon benefit weight); if the actual carbon cost is higher than expected, increase the weight. (Carbon cost weighting).

[0240] In practice, step S52, data preprocessing and labeling, is as follows:

[0241] Carbon Dioxide Benefits Before Cleaning ;

[0242] Clean carbon net benefit ;

[0243] (Estimation during the decision-making stage) ).

[0244] Actual clean carbon emissions when calculating carbon cost bias Estimated clean carbon emissions =1.06 ; (Actual carbon costs are lower than expected);

[0245] Calculate the actual reward value as the initial weight. ; .

[0246] In step S53, the target value is calculated. At that time, discount factor Target network prediction ; ;

[0247] When calculating the loss function and updating the model parameters, the loss value Mini-batch gradient descent (batch size = 32) is used, with a learning rate of... .

[0248] The weight optimization coefficients are then: ; ;

[0249] (Water consumption deviation is small, no adjustment is needed).

[0250] Example 2

[0251] See Figure 2As shown, the present invention is a photovoltaic building adaptive cleaning system based on real-time carbon accounting, which can be used to execute the method content of Embodiment 1 of the present invention, including: a data perception layer, a carbon accounting and decision-making layer and an intelligent execution layer. The three-layer structure sequentially completes data acquisition, carbon accounting and decision generation, and cleaning execution and feedback optimization.

[0252] The data sensing layer is deployed on the photovoltaic building facade and includes a power generation monitoring unit, an environmental sensor group, a surface contamination sensor, and a visual acquisition unit. The power generation monitoring unit includes a current and voltage sensor and a power meter. The environmental sensor group includes a light intensity sensor, a temperature sensor, a humidity sensor, and a wind speed sensor. The surface contamination sensor is an optical transmittance monitor with a measurement accuracy of ±0.5%. The visual acquisition unit is at least one wide-angle high-definition camera with low-light imaging capability.

[0253] The carbon accounting and decision-making layer includes a data processing module, a dynamic carbon accounting module, and an adaptive clean decision-making module. The dynamic carbon accounting module has a built-in updatable life cycle carbon emission factor library and connects to the regional power grid's real-time / predictive marginal emission factor data platform through an API interface to calculate the real-time net carbon emission reduction benefits of the photovoltaic system. The adaptive clean decision-making module integrates a deep Q-network reinforcement learning model and outputs clean decision instructions with the long-term goal of maximizing net carbon emission reduction benefits.

[0254] The intelligent execution layer includes a cleaning execution device and a drive controller; the cleaning execution device is a low-water-consumption or water-free device, selected from one or more of the following: a track-type cleaning robot, a drone cleaning system, and a local spot spraying device; the drive controller receives instructions from the decision layer and controls the cleaning execution device to perform fixed-point or area cleaning according to the planned path.

[0255] It is worth noting that the various units included in the above system embodiments are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of the present invention.

[0256] Furthermore, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware, and the corresponding program can be stored in a computer-readable storage medium.

[0257] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A photovoltaic building adaptive cleaning method based on real-time carbon accounting, characterized in that, Includes the following steps: Step S1: Collect photovoltaic system operation data, environmental data, surface contamination data, and visual image data at a frequency of no less than 5 minutes per time through the data sensing layer; the operation data includes voltage, current, and power; the environmental data includes light intensity, temperature, humidity, wind speed, and weather forecast data for the next 24 hours; the contamination data includes photovoltaic surface transmittance and dust deposition density; Step S2: Calculate the real-time net carbon reduction benefits of the system based on the dynamic carbon accounting model. The calculation logic includes the calculation of avoided carbon emissions, the separate calculation of process carbon emissions, and the coupled calculation of net carbon benefits. Step S3: Input net carbon reduction benefits, pollution data, and meteorological forecast data into the depth. Network reinforcement learning model, the model is based on the state space and reward function Output cleaning decision instructions; Step S4: If the decision instruction is to perform cleaning, the drive controller improves... The algorithm plans the cleaning path with the lowest carbon consumption and controls the cleaning execution device to complete the cleaning of fixed points or areas. Step S5: Collect power generation improvement data and carbon cost data after cleaning, and feed the data back to the adaptive cleaning decision module for in-depth analysis. Online fine-tuning and optimization of the network model forms a closed-loop operation and maintenance system.

2. The photovoltaic building adaptive cleaning method based on real-time carbon accounting according to claim 1, characterized in that, In step S1, the photovoltaic system operation data, environmental data, surface pollution data, and visual image data are calibrated and fused to output a dataset. The specific implementation process is as follows: Step S11, Equipment Deployment: Deploy the power generation monitoring unit, environmental sensor group, surface contamination sensor and visual acquisition unit, and calibrate the parameters of the equipment at the same time; Step S12, Operational Data Acquisition: The current and voltage sensors collect the voltage and loop current at both ends of the component in real time, calculate the instantaneous power based on the real-time voltage and loop current, and denoise the voltage, loop current and instantaneous power, taking the average value as the operational data for that period. Step S13: Environmental data acquisition: The light intensity sensor collects the actual irradiance on the surface of the photovoltaic module, the temperature and humidity sensor collects the surface temperature and relative humidity of the module, the wind speed sensor collects the horizontal wind speed and records the wind direction at the same time, and calls a third-party meteorological platform through the API interface to obtain the meteorological forecast data for the next 24 hours, including the probability of rainfall, wind speed range, and irradiance forecast value. Step S14, Surface contamination data acquisition: The optical transmittance monitor emits a beam of light of a specific wavelength, which is detected by the receiver after penetrating the surface of the photovoltaic module. The current transmittance is calculated and compared with the initial transmittance of the module to obtain the transmittance attenuation rate. The current dust deposition density is calculated by mapping the transmittance attenuation rate with the dust deposition density. Step S15: Visual image data acquisition: A wide-angle high-definition camera captures images of the photovoltaic facade at a preset angle, and the image data is transmitted to the data processing module for preprocessing. Step S16, Data Transmission: Transmit standardized data to the carbon accounting and decision-making layer via industrial Ethernet or 5G.

3. The photovoltaic building adaptive cleaning method based on real-time carbon accounting according to claim 1, characterized in that, In step S2, the calculation logic of the dynamic carbon accounting model includes: Avoid carbon emissions : In the formula, for Photovoltaic power generation during the period for The regional grid marginal emission factor for the time period is obtained from the power dispatch data platform via API. Process carbon emissions : In the formula, ; Water consumption for cleaning Carbon emissions from tap water production and transportation, based on data released by the local water company. The water pump consumes electricity for cleaning; Net carbon emission reduction benefits : In the formula, This refers to the system's lifecycle.

4. The photovoltaic building adaptive cleaning method based on real-time carbon accounting according to claim 1, characterized in that, In step S3, depth Network reinforcement learning models are based on the state space and reward function The specific process for issuing cleaning decision instructions is as follows: Step S31: Input net carbon emission reduction benefits Photovoltaic surface transmittance Probability of rainfall in the next 24 hours Wind speed in the next 24 hours Current power grid carbon intensity Rainwater tank water level Preprocessing operations are performed. Step S32: Encode the preprocessed data into a 5-dimensional state vector that can be recognized by the DQN model. : ； Step S33: The DQN model adopts a "convolutional layer + fully connected layer" structure, with a 5-dimensional state vector as input. The output is a 3D motion value. value; Step S34: The DQN model is iteratively updated using the Bellman equation. The value is calculated using the following formula: ； In the formula, Current state Next action The value of the action, As an immediate reward for the current action, As a discount factor, For the next state The maximum Q value for all actions; Step S35: Set the exploration rate If random number If the random number is selected, then the action is randomly selected; if the random number is selected, then the action is randomly selected. Then choose The biggest movement; Step S36: Output a clear decision instruction based on the selected action.

5. The photovoltaic building adaptive cleaning method based on real-time carbon accounting according to claim 4, characterized in that, The reward function The specific formula is: ； In the formula, This represents the expected increase in net carbon emissions after the cleanup action. , and These are the weighting coefficients.

6. The photovoltaic building adaptive cleaning method based on real-time carbon accounting according to claim 1, characterized in that, In step S4, by improving The specific process of the algorithm to plan the clean path with the lowest carbon consumption is as follows: Step S41: Construct a photovoltaic facade grid coordinate system: Divide the photovoltaic building facade into two-dimensional grid nodes according to the working units of the cleaning device, and define the coordinates of each node as follows: ;in, The coordinates are horizontal, and their values ​​range from the total horizontal length of the photovoltaic facade. The coordinates are in the vertical direction and the range of values ​​is the total vertical height of the photovoltaic facade; adjacent nodes are defined as reachable nodes, and non-adjacent nodes need to be jumped through intermediate nodes. At the same time, nodes with building structure obstacles are marked as unreachable nodes. Step S42: Calculate carbon consumption weights between nodes: Assign carbon consumption weights to the paths between every two reachable nodes. The weight calculation formula is as follows: ； In the formula, The carbon consumption weight is used for node-to-node carbon consumption. For cleaning devices from nodes To the node Mobile power consumption for The marginal emission factor of the power grid during the time period Let i be the cleaning water consumption of the area covered by the path from node i to j. Carbon emission factors for tap water / rainwater; Step S43, Improvement Algorithm initialization: Set the start and end nodes, and initialize the parameters; Step S44, Improvement The algorithm iteratively solves for the optimal path: iterates through all reachable adjacent nodes, calculates the carbon consumption of the path, selects the path with the minimum carbon consumption, and generates the optimal path; Step S45, Cleaning Execution: The drive controller converts the optimal path into motion commands for the cleaning device. Core operation nodes perform deep cleaning, while routine operation nodes perform routine cleaning.

7. A photovoltaic building adaptive cleaning method based on real-time carbon accounting according to claim 6, characterized in that, In step S44, the improvement The specific process of the algorithm iteratively finding the optimal path is as follows: Step S441: Select the node with the lowest carbon consumption among the unvisited nodes. ; Step S442: Transfer the node Add to visited collection ,from Remove from; Step S443: Traverse All reachable adjacent nodes ; Step S444: Calculate from start to Then Carbon consumption; Step S445: If the new carbon consumption is less than the current ,renew and predecessor nodes; Step S446: Backtrack the predecessor node to generate the optimal path.

8. The photovoltaic building adaptive cleaning method based on real-time carbon accounting according to claim 1, characterized in that, In step S5, the depth The specific process for online fine-tuning and optimization of the network model is as follows: Step S51, Cleaning effect and carbon cost data collection: High-frequency collection of photovoltaic operation data and resource consumption data within 24 hours after cleaning; Step S52, Data Preprocessing and Labeling: Calculate the actual carbon benefit improvement value and carbon cost deviation, and remove outliers; Step S53, Online fine-tuning of the DQN model: Update model weights based on labeled data and optimize the weight coefficients of the reward function; Step S54, Fine-tuning and solidifying parameters: Verify the decision-making effect of the new parameters; if they meet the threshold, solidify the parameters. Step S55, Closed-loop verification: Compare the carbon benefits before and after fine-tuning to confirm the optimization effect.

9. A photovoltaic building adaptive cleaning method based on real-time carbon accounting according to claim 8, characterized in that, In step S53, the loss function of the DQN model is calculated as follows: ； In the formula, The target value is calculated using the following formula: ; For the current model, the state ,action Valuation of value; For the target network parameters, Discount factor; The formula for optimizing the weighting coefficients is: In the formula, This is the first weighting coefficient after optimization. This is the learning rate.

10. A photovoltaic building adaptive cleaning system based on real-time carbon accounting, comprising a data sensing layer, a carbon accounting and decision-making layer, and an intelligent execution layer, wherein the three layers sequentially complete data acquisition, carbon accounting and decision generation, and cleaning execution and feedback optimization, characterized in that: The data sensing layer is deployed on the photovoltaic building facade and includes a power generation monitoring unit, an environmental sensor group, a surface contamination sensor, and a visual acquisition unit. The power generation monitoring unit includes a current and voltage sensor and a power meter. The environmental sensor group includes a light intensity sensor, a temperature sensor, a humidity sensor, and a wind speed sensor. The surface contamination sensor is an optical transmittance monitor with a measurement accuracy of ±0.5%. The visual acquisition unit is at least one wide-angle high-definition camera with low-light imaging capability. The carbon accounting and decision-making layer includes a data processing module, a dynamic carbon accounting module, and an adaptive clean decision-making module. The dynamic carbon accounting module has an updatable life cycle carbon emission factor library built-in and connects to the regional power grid real-time / predictive marginal emission factor data platform through an API interface to calculate the real-time net carbon emission reduction benefits of the photovoltaic system. The adaptive clean decision-making module integrates a deep Q-network reinforcement learning model and outputs clean decision instructions with the long-term goal of maximizing net carbon emission reduction benefits. The intelligent execution layer includes a cleaning execution device and a drive controller; the cleaning execution device is a low-water-consumption or water-free device, selected from one or more of the following: a track-type cleaning robot, a drone cleaning system, and a local spot spraying device; the drive controller receives instructions from the decision layer and controls the cleaning execution device to perform fixed-point or area cleaning according to the planned path.

Images