We present a comprehensive study on intelligent cleaning window prediction systems for large-scale solar panels deployed in the arid and dusty regions of Northwest China. The unique geographical and climatic conditions—abundant solar radiation, scarce rainfall, frequent sandstorms, and limited water resources—pose significant challenges to the operational efficiency and longevity of solar panels. Dust accumulation on solar panels not only reduces power generation efficiency by 8%–12% monthly on average but also induces micro-cracks and hotspots due to uneven heating. Traditional fixed-interval cleaning strategies (e.g., once per month) fail to capture the dynamic nature of dust deposition and weather events, leading to unnecessary resource consumption and missed optimal cleaning opportunities. To address these issues, we propose an intelligent prediction system that integrates multi-source real-time data, machine learning algorithms, and grid-based management to accurately forecast the best cleaning windows for solar panels. This system aims to minimize cleaning frequency, conserve water and labor, maximize energy yield, and extend the service life of solar panels. We detail the technical requirements, key innovations, and performance indicators of the system, supported by experimental data and theoretical models. The research is funded by the 2025 Wuwei Municipal Science and Technology Plan Self-raised Project (Project No. WW25Z01GY004).
Introduction
Solar photovoltaic (PV) power generation has been widely adopted as a clean and renewable energy technology. However, with the large-scale deployment of solar panels, the problem of dust and contamination on the panel surfaces has become increasingly prominent, negatively affecting power generation efficiency and overall system performance. Regular cleaning is essential to maintain and improve the performance of solar panels. Nevertheless, existing cleaning approaches are typically based on experience and fixed rules, failing to account for the actual environmental variations and the real-time status of the solar panels. This lack of scientific cleaning management leads to waste of manpower, financial resources, and water. In the northwestern arid regions of China, characterized by abundant sunlight, low rainfall, and frequent sandstorms, the challenge is even more severe. Solar panels in these areas experience dust deposition rates exceeding 2 g/m² per day—three times the national average—and monthly efficiency drops of 8%–12%. Water scarcity further compounds the issue: traditional high-pressure water cleaning consumes 30–50 tons per megawatt per cleaning, while the average water resource per mu (0.067 hectares) in the northwest is only one-fourth of the national average. Intelligent prediction systems can reduce cleaning frequency by 20%–40%, resulting in water savings of 0.8–1.2 million yuan per year per megawatt. In this context, we propose a research framework for an intelligent cleaning window prediction system tailored to the dust climate of Northwest China.
State of the Art in Cleaning Window Prediction
Domestic Technology Trends
With the advancement of artificial intelligence, Chinese government policies have increasingly supported intelligent cleaning of solar panels. The “dual carbon” goals drive the development of clean energy technologies, with key players like Huawei and Sungrow Power participating actively, alongside academic institutions. Between 2018 and 2023, cleaning strategies were dominated by empirical models based on fixed cycles (e.g., monthly cleaning) or simple weather thresholds (e.g., cleaning after rainfall). Since 2023, companies have introduced intelligent monitoring systems integrating light and dust sensors, combined with LSTM algorithms, IoT, and machine learning to predict cleaning cycles. State Power Investment Corporation has experimented with multimodal data fusion, incorporating satellite remote sensing (e.g., Fengyun-4 meteorological data), drone inspection images, and ground sensor data to improve prediction accuracy.
International Technology Trends
European and American countries are generally more advanced in foundational algorithms and cross-disciplinary integration. They utilize more complex machine learning models combined with satellite data. For instance, the United States employs fine-grained physical modeling based on dust adhesion models considering the surface charge distribution of solar panels, coupled with meteorological data to predict cleaning windows with errors less than 6 hours. European nations emphasize environmental protection, often adopting waterless cleaning technologies or robotic systems. Germany has developed self-powered roadmobile robots that integrate visual recognition and robotic arms to achieve a fully closed loop of prediction, localization, and cleaning. Japan, with limited land, focuses on efficient cleaning; for example, the University of Tokyo has developed superhydrophobic coatings for solar panels, combined with AI prediction, reducing cleaning frequency to 1–2 times per year. In the Middle East, where sandstorms are prevalent, Israeli solutions combine electrostatic dust removal modules with prediction systems, achieving zero-water cleaning in desert environments.
Technical Requirements Analysis
The northwestern arid region of China presents both opportunities and challenges for solar PV. Abundant sunlight is offset by rapid dust accumulation, high cleaning demand, and severe water scarcity. Traditional cleaning methods are costly and environmentally unfriendly. Therefore, intelligent prediction technology has significant market demand. The region serves as a core testing ground due to its high dust environment, strong policy support, and low electricity price sensitivity, shifting the demand from “optional” to “essential.”
Regional Environmental Drivers
Northwest China experiences 15–30 sandstorm days per year. In the Hexi Corridor and surrounding areas, the daily dust deposition on solar panels exceeds 2 g/m², leading to an average monthly efficiency drop of 8%–12%. Fixed-interval cleaning (e.g., monthly) fails to respond promptly after sandstorms, causing single-event efficiency losses exceeding 15%. Water resources are extremely limited: the average water availability per mu in the northwest is only one-quarter of the national average. Traditional high-pressure water cleaning consumes 30–50 tons per megawatt per cleaning. Intelligent prediction systems can reduce cleaning frequency by 20%–40%, yielding direct water-saving economic benefits of 0.8–1.2 million yuan per year per megawatt.
Policy and Subsidy Incentives
The national “14th Five-Year Plan for Renewable Energy Development” requires the promotion of intelligent operation and maintenance technologies in northwestern bases. Since 2023, six provinces including Gansu and Ningxia have offered an additional subsidy of 0.015 yuan/kWh for PV stations adopting AI-based cleaning systems.
Research Content and Key Technologies
Overall System Architecture
The primary goal of this system is to achieve intelligent prediction of cleaning timing. By applying machine learning algorithms and fusing multi-source real-time data, we can accurately predict the optimal cleaning windows for solar panels before dust accumulation becomes severe. Fine-grained management divides the PV field into multiple management grids to enable detailed panel management. A decision support system provides cleaning window recommendations and schemes for different management levels. This allows operators to better plan and execute cleaning tasks, reduce manual intervention, and improve management efficiency. The system design includes the following main modules: a multi-source data real-time collection system, a cross-source data fusion prediction model integration system, and a grid-based management decision support system.
The multi-source data collection system gathers real-time data on dust shading rate, power generation performance, and local meteorological conditions via various data interfaces. These data are stored in a database for use by the cross-source fusion prediction model and the grid-based decision support system. The cross-source fusion system preprocesses data, develops machine learning models for different functions, and integrates them into a prediction library. The decision support system covers grid management architecture design, data integration, and user interface development, calling data from the database and algorithms to display optimal cleaning windows. System integration and testing ensure practical application, with regular maintenance and optimization to adapt to changing environmental conditions.
Key Technology 1: Multi-Source Real-Time Data Collection System
The real-time data collection module is the first step, based on multi-source data acquisition and integration. We use existing dust shading sensors that detect dust accumulation on solar panel surfaces. Additionally, data from weather prediction systems are integrated to consider future weather conditions affecting cleaning timing. Power generation data (e.g., inverter data, energy metering) provide information on panel performance. This module ensures timeliness and accuracy of data for subsequent analysis. The following table summarizes the main performance indicators of the multi-source data collection system.
| Parameter | Requirement | Technical Description |
|---|---|---|
| Dust sensor | Detection accuracy ≤ 0.1 g/m², resolution 0.01 g/m² | Laser scattering principle, capable of identifying salt-alkali dust composition (e.g., high pH sand dust in northwest) |
| Meteorological monitoring module | Wind speed range 0–30 m/s, temperature –40°C to +85°C | Wide-temperature MEMS sensor, IP68 protection, suitable for extreme sandstorm environment |
| Edge computing terminal | Computing power ≥ 4 TOPS, support LoRa/5G dual-mode communication | Deploy lightweight AI model for real-time processing of 1 km² grid meteorological data |
The dust sensor must be calibrated to ensure accurate monitoring. Weather data acquisition interfaces are established to retrieve real-time weather conditions, radiation levels, etc. Power generation data interfaces collect inverter and energy meter data. A data storage platform or data warehouse is built to store and manage multi-source data securely, with backup for other modules.
Key Technology 2: Cross-Source Data Fusion Optimization Algorithm Integration System
This system incorporates different machine learning algorithms closely integrated into the prediction system to provide real-time, personalized recommendations. These algorithms are trained on historical data, including dust accumulation, cleaning timing, weather conditions, etc., to predict the optimal cleaning window and classify management grids. The algorithms fuse multiple data sources for comprehensive assessment. The following table shows the main performance indicators.
| Parameter | Requirement | Technical Description |
|---|---|---|
| Sandstorm prediction lead time | 6–12 hours early warning, time error ≤ 2 hours | Fuse meteorological model with local sandstorm path historical data |
| Cleaning window identification accuracy | ≥ 92% (based on historical data validation) | Use LSTM neural network with input features: dust accumulation, wind speed, irradiance, module temperature, etc. |
| False alarm rate | ≤ 5% | Introduce confidence threshold mechanism to prevent over-cleaning |
| Machine response delay | ≤ 15 minutes (from command to start) | Edge-cloud collaborative control, support dry brushing or micro-water cleaning mode |
| Power generation efficiency recovery | ≥ 95% within 24 hours after cleaning | Verify by solar panel IV curve detection |
| System reliability | Mean time between failures ≥ 1000 hours | Redundant design of key components, support remote system upgrade |
Data preprocessing and fusion are performed on the multi-source data. Machine learning algorithms are developed according to functional requirements, trained on fused data, and iterated. The trained models are integrated into a prediction library for the decision support system to call.
Key Technology 3: Grid-Based Management Decision Support System
The decision support system covers grid management architecture design, management data integration, and user interface development. It accesses the databases and algorithm integration system to display optimal cleaning windows for different sub-areas. The grid-based approach divides the entire PV field into smaller management units (e.g., 1 km² grids). Each grid has its own real-time data and prediction model, allowing localized cleaning schedules. This granularity improves overall efficiency and reduces resource waste. The system provides visual dashboards for operators to plan cleaning tasks, trigger automated cleaning commands, and generate reports.
Mathematical Models and Formulas
To quantify dust accumulation and cleaning effectiveness, we introduce several mathematical models. The dust deposition rate on solar panels can be expressed as:
$$
\dot{m}_d = k \cdot C_{\text{dust}} \cdot (1 – e^{-\alpha t})
$$
where \(\dot{m}_d\) is the dust mass per unit area per unit time, \(k\) is the deposition coefficient dependent on wind speed and humidity, \(C_{\text{dust}}\) is the ambient dust concentration, \(\alpha\) is the adhesion coefficient, and \(t\) is the exposure time.
The power generation efficiency loss due to dust is approximated by:
$$
\eta_{\text{loss}} = 1 – \frac{P_{\text{actual}}}{P_{\text{clean}}} = \beta \cdot m_d
$$
with \(\beta\) being the efficiency loss coefficient (typically 0.04–0.06 per g/m² for common dust types). The cleaning window is determined by minimizing the cost function:
$$
J(t) = \int_{t_{\text{last}}}^{t} C_{\text{loss}}(\tau) d\tau + C_{\text{clean}}(t)
$$
where \(C_{\text{loss}}\) is the revenue loss per unit time due to reduced power output, and \(C_{\text{clean}}(t)\) is the cleaning cost which is a step function of time (e.g., lower cost if cleaning coincides with low-price electricity periods). The optimal cleaning time \(t^*\) satisfies:
$$
\frac{dJ(t)}{dt} = C_{\text{loss}}(t^*) + \frac{dC_{\text{clean}}(t^*)}{dt} = 0
$$
Using a machine learning model such as LSTM, we predict the future trajectory of dust accumulation and thus the optimal cleaning time. The LSTM prediction error is defined as:
$$
\text{RMSE} = \sqrt{\frac{1}{N}\sum_{i=1}^{N}(y_i – \hat{y}_i)^2}
$$
where \(y_i\) is the actual dust level and \(\hat{y}_i\) is the predicted dust level. With our proposed system, we achieve RMSE values below 0.05 g/m² for a 6-hour prediction horizon.
Experimental Validation and Expected Performance
We conducted preliminary experiments using historical data from a 50 MW PV station in the Hexi Corridor. The system prototype achieved sandstorm prediction lead time of 8 hours with mean error ±1.5 hours. The cleaning window identification accuracy reached 93.2% based on 90-day historical data, with a false alarm rate of 4.3%. The edge computing terminal processed 1 km² grid data within 2 seconds. After implementing the intelligent cleaning schedule, the average cleaning frequency reduced from 18 times per year to 12 times per year, saving approximately 40% of water consumption while maintaining efficiency recovery above 96%. These results confirm the effectiveness of the proposed approach.
Conclusion
We have presented a comprehensive study on intelligent cleaning window prediction for solar panels under the dusty climate of Northwest China. The system integrates multi-source real-time data, cross-source data fusion machine learning models, and grid-based decision support to accurately determine the optimal cleaning timing. The key innovations include: (1) a dust-specific sensor network with high accuracy and robustness, (2) an LSTM-based fusion algorithm that combines meteorological, dust, and power generation data, (3) a hierarchical grid management framework that enables localized cleaning recommendations, and (4) a decision support system that minimizes resource waste. The research aligns with national policies on smart energy and water conservation. The modular, scalable design allows easy replication to other regions with similar dust challenges. Future work will focus on integrating satellite remote sensing data and exploring advanced waterless cleaning technologies to further enhance sustainability.