An intelligent photovoltaic panel health assessment and dynamic cleaning decision system and method

Abstract

The present application relates to the technical field of intelligent operation and maintenance of photovoltaic power generation systems, and specifically relates to a system and method for intelligent photovoltaic panel health degree evaluation and dynamic cleaning decision-making. The system for intelligent photovoltaic panel health degree evaluation and dynamic cleaning decision-making comprises: a data acquisition module for acquiring visual image data, real-time power generation performance data, environmental meteorological data and historical operation and maintenance data of the photovoltaic panel to be evaluated; a health degree evaluation module having a first deep learning model deployed thereon, the first deep learning model being configured to fuse the visual image data, the real-time power generation performance data and the environmental meteorological data, and output a health degree score continuously representing the current power generation capacity of the photovoltaic panel; and a dynamic decision-making module having a second deep learning model deployed thereon, the second deep learning model being configured to receive the health degree score, future meteorological prediction data and cost parameters of different cleaning modes, and dynamically generate a cleaning decision instruction including a cleaning time and a cleaning mode, with the maximum expected net income serving as an objective function.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for photovoltaic power generation systems, specifically to an intelligent photovoltaic panel health assessment and dynamic cleaning decision-making system and method. Background Technology

[0002] Surface contamination of photovoltaic panels is one of the core factors leading to a significant decrease in their power generation efficiency. To address this issue, existing technologies primarily focus on improving the cleaning execution end, such as employing automated cleaning devices and various cleaning methods. While these methods achieve some cleaning effect, systemic deficiencies remain in the cleaning triggering and decision-making mechanisms. Current mainstream solutions still rely on fixed cycles or human experience, failing to detect the dynamic accumulation of contaminants. Although some intelligent systems have begun to introduce sensors for stain identification and path planning, their decision-making logic is often limited to the physical characteristics of the stains (such as type, hardness, and distribution) or a single path optimization objective, failing to construct a decision-making model aimed at maximizing the overall economic benefits of the power plant.

[0003] In assessing pollution status, current methods rely on single indicators such as power generation deviation, which are highly susceptible to interference from factors such as weather and component aging. This results in inaccurate assessment methods and a tendency to make misjudgments. Although some technologies have attempted to combine data from multiple sensors, such as vision and ultrasound, they have failed to deeply integrate multimodal information such as power generation performance, environmental meteorology, and high-value visual images (e.g., thermal imaging). Consequently, the assessment of the "health" of photovoltaic panels is incomplete and has limited accuracy.

[0004] Furthermore, existing automation solutions for cleaning execution suffer from reliability and adaptability issues. Cleaning robots face challenges such as component wear, inadequate path planning, insufficient adaptability to complex terrain, and communication failures, affecting the stability and coverage of cleaning results. Moreover, while non-contact technologies such as ultrasonic dust removal can avoid wear and water consumption to some extent, they require structural modifications to photovoltaic modules, resulting in high costs and making them difficult to promote and apply in existing power plants. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent photovoltaic panel health assessment and dynamic cleaning decision system to solve the problems of high misjudgment rate and lack of economic optimization in the existing technology.

[0006] To address the aforementioned problems, this invention proposes an intelligent photovoltaic panel health assessment and dynamic cleaning decision-making system. The technical solution adopted is as follows: A smart photovoltaic panel health assessment and dynamic cleaning decision system includes: a data acquisition module for acquiring visual image data, real-time power generation performance data, real-time environmental meteorological data, future meteorological forecast data, and historical operation and maintenance data of the photovoltaic panel to be assessed; A health assessment module is connected to the data acquisition module and has a first deep learning model deployed therein. The first deep learning model is configured to fuse visual image data, real-time power generation performance data and environmental meteorological data, and output a health score that continuously represents the current power generation capacity of the photovoltaic panel. The dynamic decision-making module is connected to the health assessment module and the data acquisition module respectively. It has a second deep learning model deployed in it. The second deep learning model is configured to receive health scores, future weather forecast data, historical operation and maintenance data and cost parameters of different cleaning methods, and dynamically generate cleaning decision instructions that include cleaning timing and cleaning methods with the objective function of maximizing expected net benefits.

[0007] Furthermore, the health assessment module includes: The multimodal input layer is used to preprocess fused visual image data, real-time power generation performance data, and environmental meteorological data to obtain heterogeneous data; The first deep learning model is a multimodal fusion neural network based on an attention mechanism, which performs weighted feature fusion on the input heterogeneous data to quantify the contribution weight of different types of data to the health score. The output layer processes the weighted feature fusion and outputs a health score that continuously represents the current power generation capacity of the photovoltaic panel. At the same time, the output layer diagnoses the fault type based on the photovoltaic panel health score.

[0008] Furthermore, the first deep learning model includes: The parallel feature extraction network includes a visual encoding sub-network, a temporal data encoding sub-network, and an environmental multi-source network. The visual encoding sub-network uses a pre-trained ResNet-34 network to extract visual features from pre-processed fused visual image data. The time-series data encoding sub-network uses a one-dimensional convolutional neural network combined with a long short-term memory network to extract time-series features from preprocessed real-time power generation performance data. The environmental multi-source network is used to extract environmental features from preprocessed real-time environmental meteorological data. The core fusion layer combines visual features, temporal features, and environmental features, and dynamically weights and fuses these features through a self-attention mechanism.

[0009] Furthermore, the operation of the health assessment module includes: The multimodal input layer preprocesses the fused visual image data, real-time power generation performance data, and environmental meteorological data to obtain normalized visual images, power generation performance time-series data, and meteorological time-series data. The first deep learning model extracts visual features, temporal features, and environmental features, and concatenates their corresponding feature vectors. The concatenation is then input into a self-attention layer or cross-attention layer in the weight calculation module. The model automatically learns and outputs the dynamic weights of different features in specific scenarios. The dynamic weights of different features in specific scenarios are normalized to obtain the attention coefficients of visual features, temporal features, and environmental features, which are then used for weighted feature fusion. The output layer maps the weighted feature fusion features to a continuous value between 0 and 100 through a fully connected regression layer, and outputs the photovoltaic panel health score. At the same time, the output layer performs fault type diagnosis based on the photovoltaic panel health score.

[0010] Furthermore, the second deep learning model is a deep reinforcement learning model. Its state space includes the current health scores of all target photovoltaic panels, future weather forecast data, cumulative number of days without cleaning, and the panel tilt angle of the photovoltaic panels. Its action space includes no cleaning, dry brush cleaning, water washing cleaning, and cleaning intensity. The reward function of the second deep learning model is R=Δ power generation revenue-cleaning cost-λ×cumulative pollution penalty, where λ is a weight coefficient.

[0011] Furthermore, the dynamic decision-making module employs a deterministic policy gradient algorithm, specifically: Input the health scores of all target photovoltaic panels at decision time t, future weather forecast data, cumulative number of days without cleaning, and panel tilt angle of the photovoltaic panels; Based on the action probability output by its internal policy network, select the action At with the highest expected cumulative reward; Specifically, if the health score is less than a preset threshold, the system will not clean the device and will maintain the current state. When the health score is greater than or equal to a preset threshold, the cost parameters of different cleaning methods are obtained based on the second deep learning model, and detailed instructions including the cleaning photovoltaic panel number, the recommended cleaning method, and the recommended execution time are generated.

[0012] Furthermore, the computation process of the second deep learning model is as follows: Obtain the cost parameters of different cleaning methods, and calculate the expected net benefit of each candidate cleaning action based on the reward function; If there is no expected net income, then there is no economically feasible solution, and no cleaning will be performed in this case. If there is an expected net benefit, the action with the highest expected net benefit is selected, and a cleaning decision instruction including the cleaning timing and cleaning method is dynamically generated.

[0013] Furthermore, the second deep learning model is trained as follows: In the initial stage, the DRL model was trained extensively offline in a simulation environment based on the photovoltaic panel pollution accumulation model, weather model, and revenue model to configure a strategy that maximizes long-term net revenue. During the online operation phase, the actual net profit is continuously fed back to the DRL model for online fine-tuning.

[0014] Furthermore, the system also includes an execution feedback module, used to control the cleaning robot to perform cleaning operations according to the cleaning decision instructions, and to re-collect visual image data and power generation performance data after the operation to verify the cleaning effect. If the health score meets the preset cleaning standard, the cleaning record is updated and fed back to the second deep learning model; If the health score does not meet the preset cleaning standard, a second cleaning instruction will be triggered.

[0015] This invention also provides a method for intelligent photovoltaic panel health assessment and dynamic cleaning decision-making, including the following steps: Step 1: The data acquisition module acquires visual image data, real-time power generation performance data, real-time environmental meteorological data, future weather forecast data, and historical operation and maintenance data of the photovoltaic panel to be evaluated; Step 2: The health assessment module receives fused visual image data, real-time power generation performance data, and environmental meteorological data through the first deep learning model deployed on-premises, and outputs a health score that continuously represents the current power generation capacity of the photovoltaic panel; Step 3: The dynamic decision-making module receives health scores, future weather forecasts, historical operation and maintenance data, and cost parameters of different cleaning methods through the internally deployed second deep learning model. With the objective function of maximizing expected net benefits, it dynamically generates cleaning decision instructions that include cleaning timing and cleaning methods.

[0016] Beneficial Effects: This invention is an improved invention. In this invention, the health assessment module employs a first deep learning model to comprehensively analyze collected data and output a health score characterizing the instantaneous power generation capacity and long-term performance degradation trend of photovoltaic panels. The dynamic decision-making module, based on a second deep learning model, uses the health score and future weather forecast as states, the costs of different cleaning methods as actions, and maximizing expected net power generation revenue as the reward function to dynamically generate the optimal cleaning strategy. This invention achieves a shift from "regular maintenance" to "on-demand, economical maintenance," significantly improving operation and maintenance efficiency and return on investment. This invention introduces a deep reinforcement learning decision-making model with the goal of maximizing net revenue, incorporating factors such as cleaning costs, power generation revenue, and future weather into closed-loop optimization, achieving a shift from "scheduled cleaning" to "on-demand, economical cleaning," significantly reducing operation and maintenance costs.

[0017] The health assessment module includes: The multimodal input layer is used to preprocess fused visual image data, real-time power generation performance data, and environmental meteorological data to obtain heterogeneous data; The first deep learning model is a multimodal fusion neural network based on an attention mechanism, which performs weighted feature fusion on the input heterogeneous data to quantify the contribution weight of different types of data to the health score. The output layer processes the weighted feature fusion and outputs a continuous health score representing the current power generation capacity of the photovoltaic panel. Simultaneously, the output layer diagnoses fault types based on the photovoltaic panel's health score. Through multimodal data fusion and an attention mechanism using a first deep learning model, it comprehensively considers multi-dimensional information such as visual perception, power generation performance, and environmental factors, effectively distinguishing efficiency decline caused by pollution from other factors and significantly improving the accuracy of state assessment.

[0018] The system also includes an execution feedback module, used to control the cleaning robot to perform cleaning tasks according to the cleaning decision instructions, and to re-collect visual image data and power generation performance data after the task to verify the cleaning effect. If the health score meets the preset cleaning standard, the cleaning record is updated and fed back to the second deep learning model; If the health score fails to meet the preset cleaning standards, a secondary cleaning instruction is triggered. The execution feedback module achieves a fully automated closed-loop process of "assessment-decision-execution-verification," improving operational efficiency and reliability. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the intelligent photovoltaic panel health assessment and dynamic cleaning decision system of the present invention; Figure 2 This is a schematic diagram of the health assessment module in the intelligent photovoltaic panel health assessment and dynamic cleaning decision system of the present invention; Figure 3 This is a schematic diagram of the core layer structure in the intelligent photovoltaic panel health assessment and dynamic cleaning decision system of the present invention; Figure 4 This is a schematic diagram of the deep reinforcement learning decision-making process of the dynamic decision-making module in the intelligent photovoltaic panel health assessment and dynamic cleaning decision-making system of the present invention. Detailed Implementation

[0020] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] The following description, with reference to the accompanying drawings, illustrates an intelligent photovoltaic panel health assessment and dynamic cleaning decision-making system and control method according to embodiments of this application.

[0023] The following is combined Figure 1 This application provides a detailed description of the intelligent photovoltaic panel health assessment and dynamic cleaning decision system.

[0024] The intelligent photovoltaic panel health assessment and dynamic cleaning decision system of this application includes: a data acquisition module, used to acquire visual image data, real-time power generation performance data, real-time environmental meteorological data, future meteorological forecast data, and historical operation and maintenance data of the photovoltaic panel to be assessed.

[0025] Specifically, the data acquisition module is responsible for obtaining raw data from the physical world and the database. This includes visual image data acquisition: high-definition cameras and infrared thermal imagers are deployed at key locations on the photovoltaic panel array. The high-definition cameras are used to capture the macroscopic distribution of contaminants such as dust, snow, and bird droppings (e.g., using 2-megapixel cameras, triggered at fixed times daily and before and after cleaning), obtaining visible light images; the infrared thermal imagers are used to detect abnormal "hot spots" caused by dirt, with a temperature resolution of no more than 0.1℃, obtaining thermal images.

[0026] Real-time power generation performance data acquisition: The output voltage, output current and power generation data of each photovoltaic panel or string are read in real time from the power plant monitoring system, string inverter or current and voltage sensors via Modbus TCP protocol, with a sampling frequency of once per minute.

[0027] Real-time environmental meteorological data acquisition: Access real-time data from the power station's meteorological station, including light intensity, ambient temperature, humidity, wind speed, and wind direction; Future weather forecast data collection: By calling third-party weather forecast APIs, we obtain weather forecast data for the next 24 to 72 hours as a basis for decision-making.

[0028] Historical Operation and Maintenance Data Collection: Static historical information such as the installation date, model, tilt angle, and last cleaning time and method of the photovoltaic panels can be queried from the power plant asset management system. This information can also be queried from the cleaning robot management system.

[0029] All the data collected above are accessed through the edge gateway for multi-source data access. After the data is aggregated, initially formatted and aligned with the timestamp, it is transmitted to the cloud server and output.

[0030] The health assessment module, connected to the data acquisition module, houses a first deep learning model. This model is configured to fuse visual image data, real-time power generation performance data, and environmental meteorological data to output a continuous health score characterizing the current power generation capacity of the photovoltaic panels. The visual image data includes high-resolution visible light images and thermal imaging images. The first deep learning model analyzes abnormal temperature areas in the thermal imaging images to assist in determining the health score. Here, the first deep learning model is deployed on a cloud server.

[0031] Specifically, such as Figure 2 As shown, the health assessment module includes: a multimodal input layer, used to preprocess fused visual image data, real-time power generation performance data, and environmental meteorological data to obtain heterogeneous data; specifically, it scales the image data to a fixed size and normalizes it; it aligns and truncates time-series data of power generation performance (such as power sequences) with time-series meteorological data (such as sunshine sequences) using a sliding window to form time segments. The first deep learning model is a multimodal fusion neural network based on an attention mechanism, which performs weighted feature fusion on the input heterogeneous data to quantify the contribution weight of different types of data to the health score. The output layer processes the features after weighted feature fusion and outputs a health score that continuously represents the current power generation capacity of the photovoltaic panel; simultaneously, the output layer diagnoses fault types based on the photovoltaic panel health score.

[0032] The first deep learning model includes a parallel feature extraction network, comprising a visual encoding sub-network, a temporal data encoding sub-network, and an environment multi-source network. The visual encoding sub-network uses a pre-trained ResNet-34 network to extract visual features from pre-processed fused visual image data. Specifically, the pre-trained ResNet-34 network is used to extract visual features from pre-processed fused visual image data, including using a pre-trained ResNet-34 as the backbone network to extract high-level feature vectors from normalized visible light and thermal imaging images.

[0033] The time-series data encoding sub-network uses a one-dimensional convolutional neural network combined with a long short-term memory network to extract time-series features from preprocessed real-time power generation performance data. Specifically, the one-dimensional convolutional neural network combined with a long short-term memory network is used to extract time-series features from preprocessed real-time power generation performance data, including: using a one-dimensional convolutional neural network (1D-CNN) combined with a long short-term memory network (LSTM) to extract spatiotemporal feature vectors of power, illumination and other time-series data from preprocessed current and voltage time-series data.

[0034] A multi-source environmental network is used to extract environmental features from preprocessed real-time environmental meteorological data.

[0035] The core fusion layer concatenates visual, temporal, and environmental features, dynamically weighting and fusing them using a self-attention mechanism. This layer can automatically learn and output the dynamic weights of different modal features in specific scenarios, achieving weighted fusion.

[0036] In one specific embodiment, such as Figure 2 and Figure 3 As shown, the operation of the health assessment module includes: the multimodal input layer preprocesses the fused visual image data, real-time power generation performance data, and environmental meteorological data to obtain normalized visual images, power generation performance time-series data, and meteorological time-series data; the first deep learning model extracts visual features, time-series features, and environmental features, and assigns their corresponding feature vectors F. v F t and F e The data is concatenated and input into a self-attention layer or cross-attention layer in the weight calculation module. This automatically learns and outputs the dynamic weights of different features in a specific scenario. The dynamic weights of different features in a specific scenario are normalized to obtain the attention coefficients α, β, and γ of visual features, temporal features, and environmental features, which are then used for weighted feature fusion. The output layer maps the weighted feature fusion into a continuous value between 0 and 100 through a fully connected regression layer, obtaining and outputting a photovoltaic panel health score. Simultaneously, the output layer diagnoses the fault type based on the photovoltaic panel health score. Here, 100 represents the maximum power generation capacity under ideal clean conditions, and 0 represents complete failure. Here, the weighted feature fusion F... fused= α×F v +β×F t +γ×F e .

[0037] In another implementation, the first deep learning model is trained and deployed using supervised training with historical operational data (power generation loss rate due to pollution). After training, the model is encapsulated as a RESTful API service for the dynamic decision-making module to call in real time.

[0038] The dynamic decision-making module, connected to both the health assessment module and the data acquisition module, houses a second deep learning model. This model is configured to receive health scores, future weather forecasts, historical maintenance data, and cost parameters for different cleaning methods. With maximizing expected net benefit as the objective function, it dynamically generates cleaning decision instructions that include the timing and method of cleaning. These instructions include detailed suggestions for cleaning methods (actions) and recommended execution times (e.g., choosing a time with weak sunlight to conserve water).

[0039] Specifically, the second deep learning model is a deep reinforcement learning model. Its state space includes the current health scores of all target photovoltaic panels, future weather forecast data, cumulative number of days without cleaning, and the panel tilt angle. Its action space includes no cleaning, dry brushing, water washing, and cleaning intensity. The reward function of the second deep learning model is R = Δ power generation revenue - cleaning cost - λ × cumulative pollution penalty. Here, λ is a weighting coefficient. Δ power generation revenue is the expected additional power generation revenue calculated based on the PHI improvement value and the electricity price model using future weather forecasts. Cleaning cost is the cost of water, electricity, robot depreciation, etc., calculated according to the selected action. Cumulative pollution penalty is a small negative reward that increases over time to prevent the model from choosing "no cleaning" for an extended period, which could lead to irreversible pollution.

[0040] In one specific implementation, the dynamic decision-making module employs a deterministic policy gradient algorithm, specifically: Input the health scores of all target photovoltaic panels at decision time t, future weather forecast data, cumulative number of days without cleaning, and panel tilt angle of the photovoltaic panels; Here, the state space is defined (input): at decision time t, the state St of the model includes: (1) the average PHI value of all target photovoltaic panels and the current system state; (2) pollution trend analysis and weather forecast vectors for the next 24 hours (sunlight, temperature, and precipitation probability); (3) the cumulative number of days without cleaning; (4) the panel tilt angle of the photovoltaic panels; Based on the action probability output by its internal policy network, the action At with the highest expected cumulative reward is selected; here, the action At of the model is a discrete choice: no cleaning, dry brush cleaning (low cost, moderate water consumption) and water washing cleaning (high cost, high cleanliness); it can also be extended to more granular actions, such as cleaning intensity; Specifically, if the health score is less than the preset threshold T, the system will not clean this time and will maintain the current state. When the health score is greater than or equal to the preset threshold T, the cost parameters of different cleaning methods are obtained based on the second deep learning model, and detailed instructions including the cleaning photovoltaic panel number, the recommended cleaning method, and the recommended execution time are generated.

[0041] The computation process of the second deep learning model is as follows: Obtain the cost parameters of different cleaning methods, and calculate the expected net benefit of each candidate cleaning action based on the reward function; If there is no expected net income, then there is no economically feasible solution, and no cleaning will be performed in this case. If there is an expected net benefit, the action with the highest expected net benefit is selected, and a cleaning decision instruction including the cleaning timing and cleaning method is dynamically generated.

[0042] The second deep learning model is trained as follows: In the initial stage, the DRL model was trained extensively offline in a simulation environment based on the photovoltaic panel pollution accumulation model, weather model, and revenue model to configure a strategy that maximizes long-term net revenue. During the online operation phase, the actual net profit is continuously fed back to the DRL model for online fine-tuning.

[0043] In another specific implementation, the intelligent photovoltaic panel health assessment and dynamic cleaning decision-making system also includes: The execution feedback module is used to control the cleaning robot to perform cleaning tasks according to the cleaning decision instructions, and to re-collect visual image data and power generation performance data after the task to verify the cleaning effect. If the health score meets the preset cleaning standard, the cleaning record is updated and fed back to the second deep learning model; If the health score does not meet the preset cleaning standard, a second cleaning instruction will be triggered.

[0044] Specifically, the execution feedback module implements the decision-making process and verifies it in a closed loop, including: Command execution: The system sends cleaning decision commands to the corresponding area's cleaning robot cluster scheduling system; After receiving instructions, the robot performs the specified cleaning task; Effect verification: After the cleaning operation is completed, the camera in the data acquisition module is triggered to re-capture visible light and thermal images of the target photovoltaic panel; at the same time, the current and voltage sensors re-acquire power generation performance data to verify the cleaning effect. Closed-loop feedback: The preprocessed visible light and thermal imaging images, along with power generation performance data, are input again into the health assessment module to calculate a new PHI value. If the new PHI value reaches the preset standard (e.g., >90), the cleaning task is completed in closed loop. If the new PHI value does not meet the standard, the system analyzes the reasons for the cleaning failure, adjusts the cleaning strategy, automatically generates a "second cleaning" instruction (which may upgrade the cleaning method, such as from dry brushing to water washing), and reschedules the robot to execute the task until the standard is met or the maximum number of retries is reached. This validation data also serves as new experience for the reinforcement learning model, used to update the model strategy.

[0045] This application also provides a method for assessing the health of intelligent photovoltaic panels and making dynamic cleanup decisions, including the following steps: Step 1: The data acquisition module acquires visual image data, real-time power generation performance data, real-time environmental meteorological data, future weather forecast data, and historical operation and maintenance data of the photovoltaic panel to be evaluated; Step 2: The health assessment module receives fused visual image data, real-time power generation performance data, and environmental meteorological data through the first deep learning model deployed on-premises, and outputs a health score that continuously represents the current power generation capacity of the photovoltaic panel; Step 3: The dynamic decision-making module receives health scores, future weather forecasts, historical operation and maintenance data, and cost parameters of different cleaning methods through the internally deployed second deep learning model. With the objective function of maximizing expected net benefits, it dynamically generates cleaning decision instructions that include cleaning timing and cleaning methods.

[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. The scope of patent protection of the present invention shall be determined by the claims. Similarly, any equivalent structural changes made based on the description and drawings of the present invention shall also be included within the scope of protection of the present invention.

Claims

1. A smart photovoltaic panel health assessment and dynamic cleaning decision-making system, characterized in that, include: The data acquisition module is used to acquire visual image data, real-time power generation performance data, real-time environmental meteorological data, future weather forecast data, and historical operation and maintenance data of the photovoltaic panels to be evaluated. A health assessment module is connected to the data acquisition module and has a first deep learning model deployed therein. The first deep learning model is configured to fuse visual image data, real-time power generation performance data and environmental meteorological data, and output a health score that continuously represents the current power generation capacity of the photovoltaic panel. The dynamic decision-making module is connected to the health assessment module and the data acquisition module respectively. It has a second deep learning model deployed in it. The second deep learning model is configured to receive health scores, future weather forecast data, historical operation and maintenance data and cost parameters of different cleaning methods, and dynamically generate cleaning decision instructions that include cleaning timing and cleaning methods with the objective function of maximizing expected net benefits.

2. The intelligent photovoltaic panel health assessment and dynamic cleaning decision-making system according to claim 1, characterized in that, The health assessment module includes: The multimodal input layer is used to preprocess fused visual image data, real-time power generation performance data, and environmental meteorological data to obtain heterogeneous data; The first deep learning model is a multimodal fusion neural network based on an attention mechanism, which performs weighted feature fusion on the input heterogeneous data to quantify the contribution weight of different types of data to the health score. The output layer processes the weighted feature fusion and outputs a health score that continuously represents the current power generation capacity of the photovoltaic panel. At the same time, the output layer diagnoses the fault type based on the photovoltaic panel health score.

3. The intelligent photovoltaic panel health assessment and dynamic cleaning decision-making system according to claim 2, characterized in that, The first deep learning model includes: The parallel feature extraction network includes a visual encoding sub-network, a temporal data encoding sub-network, and an environmental multi-source network. The visual encoding sub-network uses a pre-trained ResNet-34 network to extract visual features from pre-processed fused visual image data. The time-series data encoding sub-network uses a one-dimensional convolutional neural network combined with a long short-term memory network to extract time-series features from preprocessed real-time power generation performance data; The environmental multi-source network is used to extract environmental features from preprocessed real-time environmental meteorological data. The core fusion layer combines visual features, temporal features, and environmental features, and dynamically weights and fuses these features through a self-attention mechanism.

4. The intelligent photovoltaic panel health assessment and dynamic cleaning decision-making system according to claim 3, characterized in that, The operation of the health assessment module includes: The multimodal input layer preprocesses the fused visual image data, real-time power generation performance data, and environmental meteorological data to obtain normalized visual images, power generation performance time-series data, and meteorological time-series data. The first deep learning model extracts visual features, temporal features, and environmental features, and concatenates their corresponding feature vectors. The concatenation is then input into a self-attention layer or cross-attention layer in the weight calculation module. The model automatically learns and outputs the dynamic weights of different features in specific scenarios. The dynamic weights of different features in specific scenarios are normalized to obtain the attention coefficients of visual features, temporal features, and environmental features, which are then used for weighted feature fusion. The output layer maps the weighted feature fusion features to a continuous value between 0 and 100 through a fully connected regression layer, and outputs the photovoltaic panel health score. At the same time, the output layer performs fault type diagnosis based on the photovoltaic panel health score.

5. The intelligent photovoltaic panel health assessment and dynamic cleaning decision-making system according to claim 1, characterized in that, The second deep learning model is a deep reinforcement learning model, whose state space includes the current health scores of all target photovoltaic panels, future weather forecast data, cumulative number of days without cleaning, and the panel tilt angle of the photovoltaic panels; Its action space includes no cleaning, dry brushing, water washing, and cleaning intensity; the reward function of the second deep learning model is R=Δ power generation revenue-cleaning cost-λ×cumulative pollution penalty, where λ is the weight coefficient.

6. The intelligent photovoltaic panel health assessment and dynamic cleaning decision-making system according to claim 5, characterized in that, The dynamic decision-making module employs a deterministic policy gradient algorithm, specifically: Input the health scores of all target photovoltaic panels at decision time t, future weather forecast data, cumulative number of days without cleaning, and panel tilt angle of the photovoltaic panels; Based on the action probability output by its internal policy network, select the action At with the highest expected cumulative reward; Specifically, if the health score is less than a preset threshold, the system will not clean the device and will maintain the current state. When the health score is greater than or equal to a preset threshold, the cost parameters of different cleaning methods are obtained based on the second deep learning model, and detailed instructions including the cleaning photovoltaic panel number, the recommended cleaning method, and the recommended execution time are generated.

7. The intelligent photovoltaic panel health assessment and dynamic cleaning decision-making system according to claim 6, characterized in that, The calculation process of the second deep learning model is as follows: Obtain the cost parameters of different cleaning methods, and calculate the expected net benefit of each candidate cleaning action based on the reward function; If there is no expected net income, then there is no economically feasible solution, and no cleaning will be performed in this case. If there is an expected net benefit, the action with the highest expected net benefit is selected, and a cleaning decision instruction including the cleaning timing and cleaning method is dynamically generated.

8. The intelligent photovoltaic panel health assessment and dynamic cleaning decision-making system according to claim 7, characterized in that, The second deep learning model is trained as follows: In the initial stage, the DRL model was trained extensively offline in a simulation environment based on the photovoltaic panel pollution accumulation model, weather model, and revenue model to configure a strategy that maximizes long-term net revenue. During the online operation phase, the actual net profit is continuously fed back to the DRL model for online fine-tuning.

9. The intelligent photovoltaic panel health assessment and dynamic cleaning decision-making system according to claim 1, characterized in that, The system also includes an execution feedback module, used to control the cleaning robot to perform cleaning tasks according to the cleaning decision instructions, and to re-collect visual image data and power generation performance data after the task to verify the cleaning effect. If the health score meets the preset cleaning standard, the cleaning record is updated and fed back to the second deep learning model; If the health score does not meet the preset cleaning standard, a second cleaning instruction will be triggered.

10. A method for intelligent photovoltaic panel health assessment and dynamic cleaning decision-making, characterized in that, The intelligent photovoltaic panel health assessment and dynamic cleaning decision system based on any one of claims 1-9 includes the following steps: Step 1: The data acquisition module acquires visual image data, real-time power generation performance data, real-time environmental meteorological data, future weather forecast data, and historical operation and maintenance data of the photovoltaic panel to be evaluated; Step 2: The health assessment module receives fused visual image data, real-time power generation performance data, and environmental meteorological data through the first deep learning model deployed on-premises, and outputs a health score that continuously represents the current power generation capacity of the photovoltaic panel; Step 3: The dynamic decision-making module receives health scores, future weather forecasts, historical operation and maintenance data, and cost parameters of different cleaning methods through the internally deployed second deep learning model. With the objective function of maximizing expected net benefits, it dynamically generates cleaning decision instructions that include cleaning timing and cleaning methods.

Images