Komorebi AI

Machine Learning nowcasting of PV Energy Using Satellite Data

Neural Processing Letters, 2019


This study focuses on photovoltaic (PV) energy production prediction in peninsular Spain using satellite-measured radiances and clear sky irradiance estimates. Four different machine learning models were applied – two linear (Lasso and linear Support Vector Regression (SVR)) and two non-linear (Deep Neural Networks, specifically Multilayer Perceptrons (MLPs), and Gaussian SVRs). The performance of these models was measured against a clear sky-based persistence model across prediction horizons up to six hours.

The Gaussian SVR outperformed other models with its errors growing slowly over time (average errors of 1.92% for the first three hours, and 2.89% for the last three). MLPs performed comparably at longer horizons (average 3.1% error) but fell short at initial ones (average 2.26% error), though still performed significantly better than the linear models. Linear models, while less accurate, provided insights into specific areas with more significant influence on PV energy predictions by leveraging the spatial sparsity offered by Lasso.

Deep Learning models to improve the predictability of wind and solar energy, thereby contributing to global sustainable energy solutions and climate change mitigation.

Where does it apply?

The predictive technology utilizing machine learning models for photovoltaic (PV) energy production can be applied across various sectors. These include renewable energy companies for optimizing energy production and distribution, government agencies for informed policy development, and research institutions for further exploration of methodologies and findings.

Moreover, they can aid weather forecasting centers in enhancing solar irradiance prediction and assist financial sectors in managing risk and formulating investment strategies in energy markets. Lastly, these models can support the refinement of smart grid management and power control strategies.

renewable energy

Why does it matters?

The reliance on photovoltaic (PV) energy, a form of renewable energy, is increasing worldwide. The integration of machine learning into its prediction processes can significantly impact the accuracy of PV energy production forecasts, contributing to more efficient energy management and grid reliability. The study findings are especially relevant as they highlight the Gaussian SVR model’s effectiveness in forecasting, which can bring improvements in energy planning.

Additionally, the spatial insights from linear models can help identify the important factors and areas influencing PV energy production, leading to informed decision-making about where future PV installations can be most beneficial. This is crucial in countries like Spain, where renewable energy sources are a growing part of their energy mix.

Machine Learning nowcasting of PV Energy Using Satellite Data

Neural Processing Letters, 2019

Machine Learning nowcasting of PV Energy Using Satellite Data

Otras publicaciones

Wind Energy Prediction
Adversarial Machine Learning: Bayesian Perspectives
Variationally Inferred Sampling through a Refined Bound

Ensemble de Árboles de Regresión para la Predicción de Energía Eólica y Radiación Solar

Aprendizaje por Refuerzo bajo Amenazas

Muestreo con Inferencia Variacional a través de un Límite Mejorado

El estudio destaca cómo los modelos de Ensemble mejoran las previsiones de energía; promoviendo una mejor gestión energética y estrategias de inversión en el sector de las energías renovables.

El estudio introduce una nueva técnica, los Procesos de Decisión de Markov bajo Amenazas, que permite mejorar la toma de decisiones en escenarios de aprendizaje por refuerzo adversario. El trabajo tiene aplicaciones en ciberseguridad, finanzas e IA.

Una nueva técnica que mejora la inferencia bayesiana en modelos probabilísticos, aumentando la eficiencia de los procesos de toma de decisiones en diversos campos. Por ejemplo, el análisis financiero, la gestión de riesgos, el aprendizaje automático y la logística.