[Paper Review] PVNet: A LRCN Architecture for Spatio-Temporal Photovoltaic PowerForecasting from Numerical Weather Prediction

Johan Mathe, Nina Miolane+2 more|arXiv (Cornell University)|Feb 4, 2019

Solar Radiation and Photovoltaics26 references17 citations

TL;DR

PVNet proposes a Long-Term Recurrent Convolutional Network (LRCN) that integrates dense spatiotemporal numerical weather predictions (NWP) and physical model outputs to forecast aggregated photovoltaic (PV) power across Germany. It achieves a 1.38% reduction in normalized root mean square error (nRMSE) compared to state-of-the-art models, while sensitivity analysis reveals irradiance and cloud cover as the most influential variables spatially.

ABSTRACT

Photovoltaic (PV) power generation has emerged as one of the lead renewable energy sources. Yet, its production is characterized by high uncertainty, being dependent on weather conditions like solar irradiance and temperature. Predicting PV production, even in the 24-hour forecast, remains a challenge and leads energy providers to left idling - often carbon emitting - plants. In this paper, we introduce a Long-Term Recurrent Convolutional Network using Numerical Weather Predictions (NWP) to predict, in turn, PV production in the 24-hour and 48-hour forecast horizons. This network architecture fully leverages both temporal and spatial weather data, sampled over the whole geographical area of interest. We train our model on an NWP dataset from the National Oceanic and Atmospheric Administration (NOAA) to predict spatially aggregated PV production in Germany. We compare its performance to the persistence model and state-of-the-art methods.

Motivation & Objective

• To address the challenge of accurate 24- and 48-hour spatio-temporal forecasting of photovoltaic (PV) power generation, which remains difficult due to high variability in solar irradiance and weather conditions.
• To develop a deep learning model that fully leverages both temporal and spatial dimensions of numerical weather prediction (NWP) data, rather than treating weather variables at isolated points.
• To improve forecasting accuracy beyond persistence and state-of-the-art models by incorporating dense spatial and temporal NWP inputs into a unified architecture.
• To enable model interpretability by visualizing the spatial influence of input variables on PV predictions through occlusion sensitivity analysis.
• To demonstrate that combining PVNet with sensitivity analysis can estimate the surface density of PV power production across geographic regions.

Proposed method

• The model employs a Long-Term Recurrent Convolutional Network (LRCN) architecture, combining Convolutional Neural Networks (CNNs) for spatial feature extraction with Recurrent Neural Networks (RNNs), specifically LSTMs, for temporal modeling.
• Input data include 1D time-series of historical PV power and 2D spatial grids of NWP variables (e.g., solar irradiance, temperature, cloud cover) and physical model outputs (e.g., clear sky irradiance) over Germany.
• The LRCN processes spatial NWP data through convolutional layers to extract spatial patterns, then feeds the resulting features into an LSTM to model temporal dependencies in the sequence.
• The model is trained on NOAA’s NWP dataset to predict spatially aggregated PV power output across Germany at 24- and 48-hour horizons.
• An occlusion sensitivity analysis is applied by masking spatial patches of input variables and measuring changes in predicted power to assess spatial influence.
• Results are visualized as heatmaps overlaying Germany’s geography to interpret the geographic impact of each input variable on PV forecasts.

Experimental results

Research questions

• RQ1 Can an LRCN architecture effectively model both the spatial and temporal structure of numerical weather prediction (NWP) data for PV power forecasting?
• RQ2 How does the inclusion of dense spatial NWP and physical model inputs improve forecasting accuracy compared to point-based or temporal-only models?
• RQ3 Which meteorological variables (e.g., irradiance, cloud cover, temperature) have the strongest spatial influence on PV power predictions across Germany?
• RQ4 Can occlusion sensitivity analysis reveal meaningful geographic patterns in the model’s attention to input variables?
• RQ5 To what extent can the combination of PVNet and sensitivity analysis estimate the spatial density of PV power production?

Key findings

• PVNet reduces the normalized root mean square error (nRMSE) by 1.38% compared to the state-of-the-art model for country-wide aggregated PV power forecasting.
• The occlusion sensitivity analysis reveals that solar irradiance has the highest spatial impact on PV predictions, followed by cloud cover, clear sky irradiance, persistence, and temperature.
• The sensitivity maps show higher model attention in southern and eastern Germany, which correlates with higher actual PV power density, validating the model’s spatial reasoning.
• The model successfully learns location-dependent relationships between input weather variables and PV output, demonstrating improved inference capability beyond pure prediction.
• The integration of spatial NWP data significantly enhances forecasting performance, confirming the value of dense spatial inputs over point-wise or sparse spatial data.
• Combining PVNet with sensitivity analysis enables the estimation of surface power density across regions, offering a new tool for energy planners to assess PV potential.