[Paper Review] Predictability of PV Power Grid Performance on Insular Sites without Weather Stations: Use of Artificial Neural Networks

Cyril Voyant, Marc Muselli+4 more|arXiv (Cornell University)|Jan 1, 2009

Solar Radiation and Photovoltaics7 references12 citations

TL;DR

This study proposes a transfer learning approach using Artificial Neural Networks (ANNs) trained on long-term global solar irradiance data from Ajaccio, Corsica, to predict solar performance at nearby sites without weather stations. By leveraging data from a single well-instrumented site (16 years of hourly/daily data), the method achieves accurate predictions for distant locations—Bastia (coastal) and Corte (mountainous)—with significantly lower errors than persistence models, especially for hourly forecasts and PV power output estimation at a 10 km distance.

ABSTRACT

The official meteorological network is poor on the island of Corsica: only three sites being about 50 km apart are equipped with pyranometers which enable measurements by hourly and daily step. These sites are Ajaccio (41°55’N and 8°48’E, seaside), Bastia (42°33’N, 9°29’E, seaside) and Corte (42°30’N, 9°15’E average altitude of 486 meters). This lack of weather station makes difficult the predictability of PV power grid performance. This work intends to study a methodology which can predict global solar irradiation using data available from another location for daily and hourly horizon. In order to achieve this prediction, we have used Artificial Neural Network which is a popular artificial intelligence technique in the forecasting domain. A simulator has been obtained using data available for the station of Ajaccio that is the only station for which we have a lot of data: 16 years from 1972 to 1987. Then we have tested the efficiency of this simulator in two places with different geographical features: Corte, a mountainous region and Bastia, a coastal region. On daily horizon, the relocation has implied fewer errors than a “naïve” prediction method based on the persistence (RMSE=1468 Vs 1383Wh/m² to Bastia and 1325 Vs 1213Wh/m² to Corte). On hourly case, the results were still satisfactory, and widely better than persistence (RMSE=138.8 Vs 109.3 Wh/m² to Bastia and 135.1 Vs 114.7 Wh/m² to Corte). The last experiment was to evaluate the accuracy of our simulator on a PV power grid localized at 10 km from the station of Ajaccio. We got errors very suitable (nRMSE=27.9%, RMSE=99.0 W.h) compared to those obtained with the persistence (nRMSE=42.2%, RMSE=149.7 W.h).

Motivation & Objective

• To address the lack of solar radiation measurements on insular sites with sparse weather station networks.
• To develop a transfer learning methodology for predicting solar irradiance at locations without historical data.
• To evaluate the performance of an ANN trained on one site (Ajaccio) for predicting irradiance at geographically distinct sites (Bastia and Corte).
• To assess the accuracy of the model in predicting actual PV power output at a nearby grid-connected system.
• To demonstrate the feasibility of using a single data-rich site to support solar energy planning across diverse microclimates.

Proposed method

• Trained a Multi-Layer Perceptron (MLP) ANN on 16 years (1972–1987) of hourly and daily global horizontal irradiance data from Ajaccio.
• Used 80% of data for training and 20% for testing, with 8 input neurons (past 8 time steps) and 1 output neuron for next-step prediction.
• Applied data stationarization by dividing irradiance by extraterrestrial irradiance and solar altitude (sin(θ)) to remove annual and daily periodicity.
• Relocated the trained ANN to predict irradiance at Corte (mountainous) and Bastia (coastal), using only the Ajaccio-trained model.
• Validated predictions against measured data at a 6.525 kW PV system located 10 km from Ajaccio using a linear PV efficiency model (ηPV = 13%).
• Evaluated performance using RMSE, nRMSE, and correlation coefficient (CC), comparing against a persistence baseline.

Experimental results

Research questions

• RQ1 Can an ANN trained on long-term irradiance data from one site predict solar irradiance accurately at another site without local measurements?
• RQ2 How does the performance of a transfer-learned ANN compare to a persistence-based forecast in daily and hourly irradiance prediction?
• RQ3 Does the model maintain accuracy across different microclimates—coastal (Bastia) and mountainous (Corte)—relative to the training site (Ajaccio)?
• RQ4 To what extent can a single ANN trained on one site predict actual PV power output at a nearby grid-connected system?
• RQ5 How does data normalization based on Ajaccio’s extremes affect prediction accuracy when relocating to sites with different irradiance ranges?

Key findings

• On the hourly horizon, the transfer-learned ANN achieved RMSE = 109.3 Wh/m² at Bastia and 114.7 Wh/m² at Corte, significantly lower than the persistence model (RMSE = 138.8 and 135.1 Wh/m², respectively).
• On the daily horizon, the ANN reduced RMSE to 1383 Wh/m² at Bastia (vs. 1468 Wh/m² for persistence) and 1213 Wh/m² at Corte (vs. 1325 Wh/m² for persistence).
• For PV power output prediction at a 10 km distant system, the nRMSE was 27.9% and RMSE was 99.0 Wh, outperforming the persistence model (nRMSE = 42.2%, RMSE = 149.7 Wh).
• The model showed strong correlation (CC > 0.84) with measured data, indicating reliable temporal pattern capture despite relocation.
• Overprediction of low irradiance and underprediction of high irradiance were observed in Bastia due to normalization using Ajaccio’s irradiance extremes, indicating a limitation in cross-site normalization.
• The results suggest hidden climatic links between Ajaccio, Corte, and Bastia, enabling reliable irradiance transfer across diverse topographies.