Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PVNet: A LRCN Architecture for Spatio-Temporal Photovoltaic PowerForecasting from Numerical Weather Prediction

Johan Mathe, Nina Miolane, Nicolas Sebastien, Jeremie Lequeux

TL;DR

This work tackles the challenge of day-ahead PV power forecasting at a country scale by introducing PVNet, a spatiotemporal LRCN that fuses dense NWP fields and physical-model outputs with historical PV data. The model employs a CNN to extract spatial features from NWP inputs and a BiLSTM to capture temporal dynamics over a sliding window of size $T=8$, predicting the horizon at $t_{forecast}=24$ h (and $48$ h) using data windows aligned to NOAA GFS predictions. Compared with persistence and prior state-of-the-art approaches, PVNet delivers notable improvements in RMSE and nRMSE for Germany-wide PV output, and an occlusion sensitivity analysis provides location-dependent insights into variable importance (irradiance, cloud cover, clear sky, persistence, temperature). The approach enhances grid integration planning by offering more accurate, interpretable, and region-specific forecasts, with guidance for extending the method to additional spatial data sources and interpolation schemes in future work.

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.

PVNet: A LRCN Architecture for Spatio-Temporal Photovoltaic PowerForecasting from Numerical Weather Prediction

TL;DR

This work tackles the challenge of day-ahead PV power forecasting at a country scale by introducing PVNet, a spatiotemporal LRCN that fuses dense NWP fields and physical-model outputs with historical PV data. The model employs a CNN to extract spatial features from NWP inputs and a BiLSTM to capture temporal dynamics over a sliding window of size , predicting the horizon at h (and h) using data windows aligned to NOAA GFS predictions. Compared with persistence and prior state-of-the-art approaches, PVNet delivers notable improvements in RMSE and nRMSE for Germany-wide PV output, and an occlusion sensitivity analysis provides location-dependent insights into variable importance (irradiance, cloud cover, clear sky, persistence, temperature). The approach enhances grid integration planning by offering more accurate, interpretable, and region-specific forecasts, with guidance for extending the method to additional spatial data sources and interpolation schemes in future work.

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.

Paper Structure

This paper contains 19 sections, 15 equations, 9 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Contributions and Outline
  4. Data Representation
  5. Photovoltaic Panel
  6. Numerical Weather Predictions
  7. Incorporation of Standard Models Predictions
  8. PVNet Model
  9. The CNN Module
  10. The LSTM Module
  11. Full Model
  12. Implementation Details
  13. Dataset
  14. Time Alignment
  15. Training and Validation Split
  16. ...and 4 more sections

Figures (9)

  • Figure 1: Equivalent Circuit of a Solar Cell.
  • Figure 2: Aggregated power over a few days
  • Figure 3: Temperature (in Kelvin) and Cloud Cover (unitless) maps over Germany on May 1st, 2016
  • Figure 4: NOAA GFS Irradiance over Germany for 7 days - one patch every 3 hours - time goes from left to right. Lighter color values are higher irradiance. High level view show alternance of days and nights.
  • Figure 5: Timeline of day ahead prediction
  • ...and 4 more figures