LSTM and K-means Algorithm for Irradiation and Temperature Clustering to Forecast Power Output of a PV Array Installed in Benin

Abstract

In this paper, we present a photovoltaic power forecasting technique based on the clustering over a large area (Benin territory), of parameters influencing the output power of a photovoltaic generator such as irradiation and temperature. We associated a time series clustering model (k-means) and a time series prediction model (LSTM). To perform the clustering, we identified a total of 287 points over the whole Beninese territory. For each point, irradiation and temperature data were downloaded from the NASA database from January 1st 2015 to December 31st 2021. From these 287 points, we arrived at 18 clusters for irradiation and 39 for temperature with a Silhouette score of at least 0.99 between elements of the same cluster. Then, using the LSTM model, we predicted irradiance and temperature for a 24-hour period with RMSE 3.66 W/m2 and 1.11 °C and regression coefficients R, 0.97 and 0.95 for irradiance and temperature respectively.