Predictability of PV power grid performance on insular sites without weather stations: use of artificial neural networks

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). 1 PRESENTATION AND ISSUE We present the results of the prediction of global radiation using Artificial Neural Networks (ANN) which are a popular artificial intelligence technique in the forecasting domain [1]. Inspired by biological neural networks, researchers in a number of scientific disciplines are designing ANNs to solve a variety of problems in decision making, optimization, control and obviously prediction [2-3]. In this context, our aim was to answer to the following question: Can we design an ANN of a site for which there is a lot of solar radiation data available and use this ANN to predict a PV power grid performance of another site? We tried to answer to this question with sites located on the island of Corsica (France). The island is characterized by a Mediterranean climate and a hilly terrain. The official meteorological network (from the French Meteorological Organization) is very poor: only three sites being about 50 km apart are equipped with pyranometers and 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). In this study, we focus on the prediction of global solar irradiation on a horizontal plane for daily and hourly horizon. These time steps have been chosen according to the electricity supplier (EDF: Electricité De France) who is interested in the estimation of the fossil fuel saving. It is very important for a remote site where electrification can be problematic [4], and for quantifying the solar potential available. Indeed, this is very important both for the power plant implementation and for sizing of PV array. Solar radiation has been measured for a long time, but even today there are many unknown characteristics of its behavior. So, it seems appropriate to develop a prediction methodology using the data available in another location in order to overcome the lack of weather station and the demand for renewable energy source on the island. 2 PHYSICAL PHENOMENA There are two approaches that allow quantifying solar radiation: the “physical modeling” based on physical process occurring in the atmosphere and influencing solar radiation [5], and the “statistic solar climatology” mainly based on time series analysis [5]. We have chosen to combine these two methods to improve the quality of prediction. In this work, we have used the physical phenomena in an attempt to overcome the seasonality of the resource. When studying the solar energy on the earth‟s surface with time series, habitually the nonstationarity perturbs the quality of prediction. Often it‟s necessary to apprehend the periodic phenomenon [6,7]. In daily case, there is seasonality with annual period, and in the hourly case there is in addition daily phenomenon. We can use the extraterrestrial global horizontal irradiance as stationarization setting. The deterministic component of the series is thus reduced, leaving more important place to the stochastic part (cloud cover). We chose a multiplicative pattern and we stationarize the time series by using the extra-terrestrial irradiation. We divide the time series by the coefficient of extraterrestrial insolation. In the hourly case, the Earth's rotation adds a new periodic component. The solar altitude angle can be easily linked to the number of photon interacting on a horizontal surface at ground level (day light). In a first approximation, it can be considered as directly proportional to the energy received. It is why we have chosen to remove the periodicity (annual and daily) dividing the time series by the coefficient of extraterrestrial insolation, but also by the sinus of the solar height. These two treatments (hourly and daily) have been used designing an ANN in order to predict the global radiation on horizontal plane. 3 SOLAR RADIATION PREDICTION FOR A SITE WITHOUT EXPERIMENTAL DATA MEASURED ANNs are capable of capturing the characteristics of any phenomenon with a good degree of accuracy. This technology often offers an alternative to traditional physical-based models, and excels at uncovering relationships in data. It is also a powerful non-linear estimator which is recommended when the functional form between input-output is unknown or it is not well understood. The main difficulties using the ANN technology are to have enough data (number of learning elements), to be sure of the data quality and to find the best architecture of ANN. Following the literature [8], we designed and optimized a Multi-Layer Perceptron (MLP) network which is the most used of the ANN architectures. Solar radiation data from 1972 to 1987 have been used for training and testing the MLP. 80% of the data available were used to train the MLP and 20% to test the MLP. We used the Matlab software to establish our network which has 8 entries, 3 hidden neurons and 1 output. These 8 entries are the 8 previous hourly measures considered (t, t-1, t-2, .. , t-7) of global solar irradiation (Wh/m2) and the output is the measure to predict for the next hour (t+1). In daily case, hours are replaced by days. We obtained good results with a prediction and a learning done on the same site: nRMSE less than 20.3% in daily horizon, 19.5% for hourly horizon and 16 years of learning. We want to use this ANN, trained successfully on site with lots of data available (Ajaccio), to estimate the insolation on a site which has no meteorological station and therefore no measure history. Many studies have tested this kind of prediction [1,9], often by establishing a regression of parameters like latitude, longitude, altitude, days of the year, etc. Their methodology doesn‟t use the time series property and therefore doesn‟t allow predicting the stochastic cloud cover but only an insolation average. As we developed above, in addition to the data of Ajaccio, we have two meteorological stations: one in Corte (mountain location) and the other one in Bastia (coastal location). In order to give a first answer to the question of the training relocation feasibility, we proposed to test both in Bastia and Corte, three models of prediction. In the case A, we use the MLP trained with data from Ajaccio (16 years from 1972 to 1987). In the case B, we use the MLP trained with data from the selected place. We have 5 years of data in Corte (from 2002 to 2006) and also 5 years in Bastia (from 1991 to 1995). And in the case C, we use a “naïve” prediction method based on the persistence. The following tables present our results for a full year (1996 for Bastia and 2007 for Corte) of global horizontal irradiance prediction by daily (Table 1) and hourly step (Table 2). Table I: Comparison of the 3 techniques of prediction for the daily forecast: horizon 1 day