5/5/2025
-
4
min
MLOps at the Edge: Predicting Solar Radiation in Photovoltaic Plants
This article takes you inside a real-world deployment where MLOps at the edge powers hourly solar radiation forecasts in a photovoltaic plant. By combining AI with real-time data flows, and cloud integration, the solution delivers higher prediction accuracy, faster response times, and optimized plant performance. Discover how Edge MLOps is transforming solar forecasting.
Introduction
In the rapidly evolving world of renewable energy, optimizing the performance of photovoltaic (PV) plants is a constant challenge, and a strategic imperative.
One of the most influential factors in this equation is solar radiation, the primary input for calculating expected energy output. Accurately forecasting solar radiation on an hourly basis unlocks new levels of operational efficiency.
Why Predict Solar Radiation?
1. Enhanced Energy Forecasting:
Accurate solar radiation predictions enable better energy output estimations, allowing operators to align generation with demand more precisely.
2. Improved Operational Efficiency:
Knowing in advance when radiation will dip or peak enables operators to schedule cleanings or maintenance tasks during low-irradiance periods, avoiding unnecessary downtime.
3. Maintenance Optimization:
Patterns in solar radiation data can flag potential performance issues early, leading to proactive maintenance and increased system longevity.
4. Better Grid Integration:
Forecasting radiation improves grid coordination, reducing overproduction risks and minimizing reliance on fossil-fuel-based energy reserves.
Data Inputs: Feeding the Model
The model is trained on weather data sourced from OpenMeteo. Key variables include:
- Temperature (2m): Impacts panel efficiency.
- Apparent Temperature: Reflects perceived heat, factoring in humidity.
- Cloud Cover (%): A major influence on radiation levels.
- Shortwave / Terrestrial / Direct Radiation: Various radiation types, including direct normal irradiance.
- Diffuse Radiation: Captures scattered sunlight data.
- Global Tilted Irradiance: The model's target variable, representing radiation on the PV panel surface.
- Seasonality: Adjusts for annual radiation patterns.
These features are carefully selected to train the model to predict global_tilted_irradiance_instant (W/m²), a direct proxy for energy production capacity.
Model Architecture: Neural Network Overview
The model is a deep neural network with:
- Input Layer: Receives real-time weather variables.
- 5 Hidden Layers: Use ReLU activation functions to learn nonlinear relationships.
- Output Layer: Returns hourly solar radiation predictions.
Training is based on historical data, optimizing for accuracy in real-world deployment scenarios. The model is developed using Tensorflow framework and will be served in the Edge Node using TFX engine.
Edge Deployment: AI on the Frontlines
To ensure low latency and high reliability, the model is deployed on an edge node located at the PV plant. This local deployment is crucial for real-time responsiveness and resilience in disconnected scenarios.
Components of the Edge Stack
To support the deployment of the solar irradiance predictor on the edge, a suite of integrated components works in harmony to manage data ingestion, processing, storage, visualization, and communication.
These building blocks form the backbone of the edge architecture, ensuring that data moves seamlessly and efficiently from source to insight while enabling robust, low-latency operations even in offline scenarios.
- OPC-UA Connector: Continuously fetches updated weather data from a local weather station.
- MQTT Broker: Acts as the communication backbone for all microservices.
- Solar Irradiance Predictor Model: Tensorflow AI Model that consumes weather data and publishes radiation forecasts.
- InfluxDB: A time-series database to store all weather and prediction data.
- MQTT-InfluxDB Ingester: Bridges MQTT and InfluxDB, ensuring seamless data flow.
- Grafana: Visualizes live and historical irradiance data for operators in a dashboard.
- Azure Connector: Transmits edge data to Azure IoT Hub for further analysis and storage.
- Alert Manager: Notifies operators of anomalies in real-time.
- ML Monitoring: Tracks model accuracy and performance, triggering retraining when needed.
Data Workflow: From Sensor to Insights
The solar radiation prediction system operates through a coordinated, multi-stage pipeline that captures, processes, stores, and analyzes weather data in near real-time. This pipeline ensures that predictions are both timely and actionable, supporting the day-to-day decisions of PV plant operators. Below is an overview of how data flows through the system, from acquisition to visualization.
- Data Collection: Real-time weather data ingested via OPC-UA protocol from the local weather station.
- Prediction: AI model forecasts next-hour solar radiation.
- Storage: Results are archived in InfluxDB.
- Cloud Transmission: Data forwarded to Azure IoT Hub for extended analytics.
- Alerting: Threshold breaches prompt real-time notifications.
- Model Monitoring: Ensures continued model accuracy over time.
- Visualization: Grafana dashboards provide intuitive real-time overviews.
Flexibility and Scalability: Endless Combinations
One of the strengths of this architecture is its modularity.
You can leverage any machine learning framework best suited for your needs, including: Edge Impulse models, frameworks compatible with Seldon ML, XGBoost, ONNX, TensorFlow, PyTorch, or Scikit-learn.
Similarly, the system is designed to send predictions and insights to virtually any cloud solution, such as: Google Cloud, AWS, Azure, Splunk, InfluxDB, MongoDB, Elastic, or Kafka.
This flexibility allows organizations to customize their infrastructure and tools without being locked into a single ecosystem, making the solution highly adaptable to diverse environments and use cases.
Results and Benefits
This edge-based MLOps approach empowers PV plant operators to:
- React swiftly to changing weather conditions.
- Optimize energy production and grid synchronization.
- Improve asset utilization and maintenance planning.
- Scale effortlessly by leveraging cloud integration.
Conclusion
This use case illustrates the power of MLOps at the edge. By combining AI, real-time data pipelines, and cloud integration, we’ve built a robust system for hourly solar radiation prediction in a PV plant. The result is improved forecasting, greater operational flexibility, and enhanced system performance.
As industrial AI moves to the edge, so must the infrastructure that supports it. The Barbara Edge AI Platform delivers the tools needed to operationalize Machine Learning across complex, distributed environments, securely, efficiently, and at scale. Whether you're just starting your AI journey or scaling from pilot to production, Barbara makes Edge MLOps a reality.
Want to learn more?
Book a demo or get in touch, one of our experts will be happy to answer all your questions.
4
min
As industries seek faster insights and real-time responsiveness, Edge MLOps (Machine Learning Operations at the Edge) is emerging as a game-changer. Unlike traditional MLOps that rely on centralized cloud infrastructure, Edge MLOps enables machine learning models to be deployed, monitored, and retrained close to the data source.
Edge MLOps
4
min
This article takes you inside a real-world deployment where MLOps at the edge powers hourly solar radiation forecasts in a photovoltaic plant. By combining AI with real-time data flows, and cloud integration, the solution delivers higher prediction accuracy, faster response times, and optimized plant performance. Discover how Edge MLOps is transforming solar forecasting.
Edge MLOps
5
min
For industrial companies embracing AI, understanding the MLOps workflow is key to turning use cases into real-world results. In this post, we show how Barbara integrates seamlessly into each stage of the workflow especially where the edge plays a critical role.
Edge MLOps
