Short-term Photovoltaic Prediction by using H ∞ Filtering and Clustering Yasuhiko Hosoda 1 and Toru Namerikawa 1 1 Department of System Design Engineering, Keio University, Kanagawa, Japan (Tel: +81-45-563-1151; E-mail: yasuhiko@nl.sd.keio.ac.jp, namerikawa@sd.keio.ac.jp) Abstract: This paper deals with prediction algorithm applying for photovoltaic (PV) systems in smart grid. This prediction is aim to predict the amount of the next day of generation using the previous data and the weather forecast which get from Japan Meteorological Agency. The procedure of prediction consists of two steps, the data processing and the unknown parameters estimation. In the data processing, our proposed method considers the characteristics of PV generation using cluster ensemble. We propose the cluster ensemble based on k-means to choose the groups with a correlation with previous data. In the unknown parameters estimation, we provide the regression model for PV generation and the unknown parameters are estimated via H ∞ filtering. The effectiveness of the proposed prediction method is demonstrated through numerical simulations. Keywords: PV, Short-term, Prediction, Smart Grid, Clustering, k-means, Estimation, H ∞ Filtering 1. INTRODUCTION In recent years, energy and environmental problem have become the hottest worldwide social problems. Therefore, the energy technologies such as smart grid, micro grid and sensor networks are interested and grow- ing all over the world. For example, it is required in the smart grid (see Fig. 1) that different power generators (photo-voltaic generator, wind farm, fuel cell, micro gas turbine, battery and so on) connected with each other and cooperative in the energetically and environmentally. At the same time, it is well known that PV system is one of the best solution for such an environmental and en- ergy problems. The problem in operational management of PV system is reverse power flow to the power sys- tem. The reverse power flow adversely affects the hole of power system. Moreover, from the point view of improv- ing the control performance of power system, generation power from PV system should be estimated as accuracy as possible. Thus, the predict amount of power in next day is essential to stably operation for PV system [1-3]. A great deal of efforts to developed application of pre- diction, regression prediction techniques for time series data like PV generation data has been studied for a long time. In machine learning, the regression using Kernel method has been developed. Support Vector Machine (SVM) is a kind of the Karnel method and they are used in PV prediction [4]. The neural networks have been in- vestigated for the PV prediction and it is easy to be ap- plied for PV forecasting [3, 5]. On the other hand, a lin- ear regression is one of the simplest method and they are used in various situations. It is expansible to treat a non- linear regression, and they have been successful in past many problems. Our proposed method employs the lin- ear regression model and in their estimation via H ∞ fil- tering which make it possible to estimate more accurately against initial errors, disturbances and noises. The main objective of this paper is to present the al- gorithm for PV system prediction which can be useful for the amount of generation in the next day. It is well known that the PV generation has a specific characteris- tics and difficult to predict accuracy. Thus, the prediction needs to consider this characteristics of PV generation. Hence, the prediction algorithm has two steps, the data processing and the unknown parameters estimation. In the data processing, our proposed method considers the characteristics of PV generation using cluster ensemble. At the same time, the PV system has been influenced by cell temperature. Hence, there have 2-dimensional data temperature and generation data. On the other hand, rep- resentative cluster ensemble k-means method is not be able to separate the data stably because their algorithm select the initial value randomly. Our proposed method also has 2-dimension, hence we modeled PV system, and treat the 2-dimensional data as 1-dimension. After that, we propose the cluster method based on k-means consid- ering initial value. In estimation, we provide the linear regression model for PV generation. The unknown pa- rameters are estimated via H ∞ filtering and simulations were held and got good results in prediction. Fig. 1 Smart Grid 2. PROBLEM FORMULATION 2.1 Modeling of PV System In this subsection, we present a general plant model for PV system as a space case. The efficiency of PV gen- eration is effected by the cell temperature [6-8]. Let us define the efficiency of power generation η d k (t d k ) in day d at time k. η d k (t d k ) = η 0 (1 − ζ (T r − t d k )) (1) SICE Annual Conference 2012 August 20-23, 2012, Akita University, Akita, Japan PR0001/12/0000- ¥ 400 ©2012 SICE 0119 -119-