Michel Piliougine

Binding statistical and machine learning models for short-term forecasting of global solar radiation

A model for short-term forecasting of continuous time series has been developed. This model binds the use of both statistical and machine learning methods for short-time forecasting of continuous time series of solar radiation. The prediction of this variable is needed for the integration of photovoltaic systems in conventional power grids. The proposed model allows us to manage not only the information in the time series, but also other important information supplied by experts. In a first stage, we propose the use of statistical models to obtain useful information about the significant information for a continuous time series and then we use this information, together with machine learning models, statistical models and expert knowledge, for short-term forecasting of continuous time series. The results obtained when the model is used for solar radiation series show its usefulness.

An intelligent memory model for short-term prediction: an application to global solar radiation data

This paper presents a machine learning model for short-term prediction. The proposed procedure is based on regression techniques and on the use of a special type of probabilistic finite automata. The model is built in two stages. In the first stage, the most significant independent variable is detected, then observations are classified according to the value of this variable and regressions are re-run separately for each Group. The significant independent variables in each group are then discretized. The PFA is built with all this information. In the second stage, the next value of the dependent variable is predicted using an algorithm for short term forecasting which is based on the information stored in the PFA. An empirical application for global solar radiation data is also presented. The predictive performance of the procedure is compared to that of classical dynamic regression and a substantial improvement is achieved with our procedure.

Support Vector Regression for Electricity Consumption Prediction in a Building in Japan

This work studies how to apply support vector machines in order to forecast the energy consumption of buildings. Usually, support vector regression is implemented using the sequential minimal optimisation algorithm. In this work, an alternative version of that algorithm is used to reduce the execution time. Several experiments were carried out taking into account data measured during one year. The weather conditions were used as independent variables and the consumed amount of electricity was considered as the parameter to predict. The model has been trained using the first six months of the dataset whereas it was validated using the following three months and tested taking into account the last three months of measurements. From obtained results, a good performance of the model is observed.