CIRED Workshop - Helsinki 14-15 June 2016 Paper 0372 Paper No 0372 Page 1 / 5 A PHOTOVOLTAIC PRODUCTION ESTIMATOR BASED ON ARTIFICIAL NEURAL NETWORKS Edoardo CORSETTI Antonio GUAGLIARDI Carlo SANDRONI Rse – Italy Rse – Italy Rse - Italy corsetti@rse-web.it guagliardi@rse-web.it Sandroni@rse-web.it ABSTRACT The ability to forecast the expected power production from renewable sources nowadays is increasingly critical because of the reliability expected from them. For instance improving the reliability of photovoltaic production forecasts in a small/medium microgrid permits to save money to supply its own loads and also to plan the participation to the ongoing services the smart distribution grid will require. In this paper we propose a method to predict photovoltaic production based on a statistical model. This type of models, compared to other ones, are easily configurable, cope well with heterogeneous plants, with different ageing devices and are able to consider diverse exogenous, well known and accidental, drawbacks. Such models is daily updated with the data arising from the monitoring until 5 days before and include the relevant variables for the photovoltaic forecasting function. INTRODUCTION In Italy the most abundant renewable energy source is solar energy which can be harnessed for commercial uses through solar plants, and which is predicted by numerous analyses to become the mostly used energy resource in the next future. For the time period 2030–2050, The European Photovoltaic Industry Association (EPIA) together with the European Renewable Energy Council (EREC) has shown the high potential of PV within the RE-thinking 2050 scenario [2]. PV is expected to become a mainstream power source in Europe by 2020 and a major power source in 2050 with an approximation of 962 GW of installed capacity. This is especially true due to the installation of grid connected PV systems [3]. As a result, new challenges have arisen in the PV industry over the past years such as: identifying and quantifying the impact of PV power on the electrical distribution grid; minimizing the influence of its fluctuations; managing uncertainties in order to guarantee a secure and reliable electrical power services; and the PV systems performance prediction. The performance of a photovoltaic (PV) installation is related to factors including the electrical parameters of its components, such as PV panels and inverters, the characteristics of the installation (tilt angle, orientation, etc.) and the meteorological conditions. The power produced by a (PV) field depends mainly on the absorbed solar irradiance. In fact, a correlation exists between the PV modules’ maximum power and the solar irradiance. Solar irradiance on a panel varies with geographic location, time, the orientation of the panel relative to the sun and weather conditions. This explains the variable, chaotic and intermittent behaviour of generated solar power. To ensure an efficient exploitation and a large penetration for such a power source, it is important to predict the amount of energy that a PV installation can generate. Once fully exploited during the PV installations’ design stage, PV production forecasting is, nowadays, a must to ensure an effective management system for the electrical distribution grid. Several studies were conducted to ensure this task [4][5]. The literature presents numerous models for PV modules which can be used to quantify the expected produced electric power [6]. Also a significant number of studies on solar irradiation modelling and forecasting have been undertaken, offering a wide range of possibilities gleaned from diverse areas of knowledge such as atmospheric physics, solar instrumentation, machine learning, forecasting theory and remote sensing in its quest for better predictive skills [7]. This paper presents a part of the research in progress that seeks to estimate and forecast PV installations’ production one day ahead for the RSE distributed generation microgrid. RSE microgrid is equipped with a PV field, an engine to produce electricity and heat, and a number of energy storage systems. The microgrid can be operated in connected or in islanded mode with respect to the distribution grid. Microgrid operation is conducted according to a daily plan production that exploit renewable power sources as far as possible to supply loads. The plan is elaborated taking into account forecasts of load requests and PV production supply. PV production supply forecast is computed by means of weather forecast and it is crucial to ensure the reliability of the PV field production. The methodology proposed here adopts artificial neural networks (ANNs) modelling techniques for the PV renewable sources. The methodology we are proposing set an estimator which given a daily weather forecast leads to define the microgrid PV power production. This estimator is articulated into three different ANNs: the first one estimates the Global Tilted Irradiance (GTI), the second one estimates the PV panel temperature and the last one estimates the PV power production. The rest of the paper is organized as follows: first it is proposed an overview of the problem addressed by the paper, than a brief introduction to neural networks is given and next experimental results gained on the RSE microgrid plant are reported. PROBLEM OVERVIEW To obtain a high quality weather and irradiance data is one of the most important steps in PV performance modeling, since the uncertainty in the irradiance data usually accounts a large amount of the total uncertainty. Historical data, concerning irradiance, power production and also weather variables, plays a fundamental role. However, when the results of the model are being used for large