Selection of Distributed Sensors for Multiple Time Series Prediction

In this paper we are proposing a methodology for the selection of a subset of sensors for a prediction system of a total energy production in a region. The network of wireless sensors is distributed close to power plants and generates multiple time series data representing energy production of each plant. Our study shows that we can estimate the total energy using significantly reduced number of sensors. To build the prediction model we have used three common forecasting techniques, support vector machines (SVMs), Multilayer Perceptron (MLP), and Multiple Regression (MR). For training and testing of the models we have used the data from year 2002 to 2004. Our data set consists of 201 attributes. First 200 represents the data from sensor stations and the additional variable is the total energy production. Based on our results MR technique gives the best model for the prediction and it outperforms the MLP and SVM. We analyzed the quality of prediction with different subsets of sensors. Based on this study we have estimated the minimum number of sensors required for the prediction model. We also estimated the optimum number of sensors that will balance the expenses of the system with the accuracy. Proposed model and performed analysis may embody crucial information for producers and consumers when planning bidding strategies in energy trading in order to maximize their benefits and utilities.